track reconstruction in dense environments and …with dense environments are produced much more...

CER

N-T

HES

IS-2

017-

004

23/0

1/20

17

T R A C K R E C O N S T R U C T I O N I N D E N S E E N V I R O N M E N T S A N D T H ES E A R C H F O R N E W P H Y S I C S I N T H E F U L LY H A D R O N I C D I B O S O N

C H A N N E L S W I T H T H E AT L A S D E T E C T O R

Dissertation

R O L A N D JA N S KY

November 2016

Faculty for Mathematics, Computer Science and PhysicsUniversity of Innsbruck

Roland Jansky:Track reconstruction in dense environments and the search for new physics in the fullyhadronic diboson channels with the ATLAS detector,November 2016

S U P E RV I S O R S:Emmerich KneringerAndreas Salzburger

A B S T R A C T

With the increase in center-of-mass energy of the LHC top

s = 13 TeV for Run 2, eventswith dense environments are produced much more abundantly. In the core of highlyenergetic hadronic jets, the average separation of charged particles is comparable tothe size of individual ATLAS inner detector elements. These dense environments maybe produced by new physics processes or objects, including massive particles that de-cay to highly boosted bosons. However, this density can create confusion within thealgorithms reconstructing charged particle trajectories (tracks), so careful optimiza-tion must be carried out to ensure that the track reconstruction performance in denseenvironments is not adversely affected. Such optimization will increase the possibil-ity of discovery of new phenomena and allow higher precision measurements of thenewly opened kinematic regime.

This work describes a series of improvements to the ATLAS offline track reconstruc-tion to enhance its performance in dense environments. The effects of these improve-ments are demonstrated using both Monte Carlo simulation and data. Using data alone,residual inefficiencies of the track reconstruction in the core of jets are quantified as afunction of the transverse momentum of the jet. The fraction of lost tracks is presentedusing the energy loss in silicon. It varies from 0.061± 0.006(stat.)± 0.014(syst.) to0.093±0.017(stat.)±0.021(syst.) between a transverse jet momentum of 200 to 400GeV and 1400 to 1600 GeV, respectively.

With this improved track reconstruction performance, and vastly smaller uncer-tainties through the data-driven measurement of the track reconstruction inefficiency,ways to reconstruct the masses of jets with higher precision become possible. It isdemonstrated that combining the strengths of both the calorimeter and the trackerinto a combined jet mass provides the most performant method currently availablein ATLAS, reducing both the resolution of the reconstructed mass of the jet and itsuncertainty.

Employing these novel reconstruction methods, a search for resonances with massesin the range 1.2 < m < 3.5 TeV in the hadronically decaying W Z , WW , or Z Z fi-nal state, is performed in 15.5 fb−1 of

ps = 13 TeV proton-proton collision data.

No significant deviations from the Standard Model background expectations are ob-served. An additional charged or neutral heavy vector boson, as predicted by the HeavyVector Triplet phenomenological Lagrangian (assuming gV = 1), decaying throughW ′→W Z (or Z ′→WW ), is excluded in the mass range 1.2–2.0 (1.3–1.7) TeV at the95% confidence level.

iii

Z U S A M M E N FA S S U N G

Mit dem Anstieg der Kollisionsenergie des LHC auf 13 TeV in Run 2 werden Ereignissemit hohen Teilchendichten um ein Vielfaches häufiger. Im Zentrum hochenergetischerhadronischer Jets ist die durchschnittliche Entfernung zwischen geladenen Teilchenvergleichbar mit den Dimensionen der einzelnen Sensoren des ATLAS Inner Detector.Diese hohen Teilchendichten werden häufig durch Signaturen möglicher neuer Physikerzeugt, zum Beispiel massive Teilchen welche in stark geboostete Bosonen zerfallen.Diese hohe Dichte kann jedoch zu Problemen in der Spurenrekonstruktion führen.Durch gezielte Veränderungen der Spurenrekonstruktion können diese Probleme um-gangen und deren Leistung verbessert werden. Solche Verbesserungen erhöhen sowohldie Chance neue Phänomene zu entdecken, als schaffen sie auch die Grundlage für de-taillierte Messungen des neu erschlossenen kinematischen Regimes.

Optimierungen der ATLAS Offline-Spurenrekonstruktion werden in dieser Arbeit er-läutert und die daraus resultierende, verbesserte Leistung anhand von sowohl Monte-Carlo Simulation als auch von Daten präsentiert. Mit einer vollständig datenbasierteMethode wird jegliche verbleibende Ineffizienz der Spurenrekonstruktion im Zentrumvon Jets als eine Funktion des Transversal-Impulses der Jets bestimmt. Der Bruchteilder Spuren welche einen Pixel-Cluster in der zweiten Pixel-Lage, erzeugt von zweiTeilchen, enthalten sollte und die bei der Rekonstruktion verloren gehen, wird mitHilfe des Energieverlustes der Teilchen im Silizium bestimmt. Er variiert zwischen0.061±0.006(stat.)±0.014(syst.) bei einem transversalen Impuls im Bereich von 200bis 400 GeV und 0.093± 0.017(stat.)± 0.021(syst.) bei einem transversalen Impulszwischen 1400 bis 1600 GeV.

Durch die verbesserte Leistung der Spurenrekonstruktion und die damit verbunde-nen erheblich reduzierten Unsicherheiten, durch die beschriebene datenbasierte Mes-sung der Ineffizienz, werden neue Methoden zur Bestimmung der Masse eines Jetsmöglich. Es wird gezeigt, dass eine Kombination der Stärken von Kalorimeter undTracker zu der derzeit leistungsfähigsten Methode zur Bestimmung der Jet-Masse inATLAS führt.

Mit diesen neuartigen Rekonstruktionsmethoden wird in 15.5 fb−1 anp

s = 13 TeVProton-Proton Kollisionsdaten nach Resonanzen mit Massen im Bereich von 1.2< m<3.5 TeV gesucht, welche in WW , W Z oder Z Z Bosonen-Paare zerfallen, die als geboost-ete Boson-Jets rekonstruiert werden. Es wird kein signifikanter Überschuss an Ereignis-sen über den laut Standardmodell erwartenden Untergrund Zählraten beobachtet.Zusätzliche Vektorbosonen W ′ (Z ′) mit einem Zerfall zu W Z (WW ), vorhergesagtdurch ein Modell des phänomenologischen Lagrangian des Heavy Vector Triplet, wer-den im Massenbereich zwischen 1.2–2.0 (1.3–1.7) TeV mit einer statistischen Sicher-heit von 95% ausgeschlossen.

iv

A C K N O W L E D G M E N T S

First, I would like to express my sincerest gratitude to Emmerich Kneringer. Withouthim, I would have never entered the field of high energy physics and without hisunconditional support this thesis would not exist. Dietmar Kuhn deserves a specialmention for being the soul of the Innsbruck ATLAS group and allowing me to flourishwithin it.

Andreas Salzburger and Markus Elsing provided the best guidance possible, pro-vided expert technical advice, while all the same giving me the free space I needed.Especially, I want to thank them for pushing me in order for this work to reach its fullpotential.

Anthony Morley and Gabriel Facini deserve my greatest gratitude. Both are shiningexamples of not just brilliant physicists, but also fantastic colleagues. Anthony enduredan innumerable amount of questions over the years, but still keeps providing both themost critical and useful answers imaginable. Gabriel enabled me to see the big picturein today’s particle physics. His passion and positivity are an inexhaustible source ofmotivation. Many more colleagues will remain unnamed, but I want to thank the wholeTIDE group, which I was very proud to work in and convene.

Let me thank Heather Gray, for always having advice when needed. James Catmore,apart from being a great friend, his unrivaled knowledge of the English language prof-ited this thesis immensely. Yanyan Gao, has my deepest appreciation for welcomingme into her analysis team and putting up with my occasionally effervescent nature.

My family, including Alexandra Feistmantl, can not be thanked enough. Thank you.

v

C O N T E N T S

1 I N T R O D U C T I O N 1

I T H E O RY & AT L A S E X P E R I M E N T 32 T H E O RY 5

2.1 The Standard Model of Particle Physics . . . . . . . . . . . . . . . . . . . . 53 T H E L H C A N D T H E AT L A S E X P E R I M E N T 7

3.1 Large Hadron Collider at CERN . . . . . . . . . . . . . . . . . . . . . . . . . 73.2 Coordinate System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83.3 The Magnet System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83.4 The Inner Detector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93.5 The Calorimeter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123.6 The Muon Spectrometer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133.7 The Trigger System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133.8 The ATLAS Computing Network . . . . . . . . . . . . . . . . . . . . . . . . . 133.9 The ATLAS Offline Software - Athena . . . . . . . . . . . . . . . . . . . . . . 13

II T R A C K R E C O N S T R U C T I O N I N D E N S E E N V I R O N M E N T S 154 T R A C K R E C O N S T R U C T I O N 17

4.1 Pixel and SCT Clusterization . . . . . . . . . . . . . . . . . . . . . . . . . . . 174.2 Iterative Combinatorial Track Finding . . . . . . . . . . . . . . . . . . . . . 174.3 Track Candidates and Ambiguity Solving . . . . . . . . . . . . . . . . . . . 184.4 Truth-Based Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204.5 Truth-Based Track Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

5 A L G O R I T H M I C I M P R O V E M E N T S T O T H E T R A C K R E C O N S T R U C T I O N 215.1 Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215.2 Merged Clusters on Reconstructed Tracks . . . . . . . . . . . . . . . . . . . 215.3 Limitations on Shared Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . 26

6 R U N 2 P E R F O R M A N C E A N D S T U D I E S W I T H T H E I T K L AY O U T 316.1 Jet Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316.2 Track Reconstruction Performance in the Core of Jets . . . . . . . . . . . . 326.3 Implications for Flavor Tagging . . . . . . . . . . . . . . . . . . . . . . . . . 346.4 Performance of Track Reconstruction in Dense Environments with the

ITk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

III M E A S U R I N G T R A C K R E C O N S T R U C T I O N P E R F O R M A N C E I N D ATA 417 P R O P E RT I E S O F R E C O N S T R U C T E D T R A C K S I N D E N S E E N V I R O N M E N T S

I N R U N 2 D ATA 438 M E A S U R I N G T H E P E R F O R M A N C E O F T H E P I X E L C L U S T E R N E U R A L N E T-

W O R K I N D ATA 478.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 478.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

vii

viii C O N T E N T S

8.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559 M E A S U R I N G T R A C K L O S S I N D E N S E E N V I R O N M E N T S W I T H T H E D E/D X

M E T H O D 579.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 579.2 Systematic Uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 619.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

IV J E T M A S S R E C O N S T R U C T I O N 6710 J E T M A S S R E C O N S T R U C T I O N 69

10.1 Jet Mass Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6910.2 Jet Mass Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7110.3 Jet Mass Performance in Simulation . . . . . . . . . . . . . . . . . . . . . . 7310.4 Jet Mass Systematic Uncertainties . . . . . . . . . . . . . . . . . . . . . . . . 7510.5 Combination of mTA and mcalo . . . . . . . . . . . . . . . . . . . . . . . . . . 7610.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

V S E A R C H F O R R E S O N A N C E S W I T H B O S O N -TA G G E D J E T S 7911 S E A R C H F O R R E S O N A N C E S W I T H B O S O N -TA G G E D J E T S 81

11.1 Data Sample and Simulated Samples . . . . . . . . . . . . . . . . . . . . . . 8211.2 Preselection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8411.3 Event Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8511.4 Control Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9511.5 Background Parametrization . . . . . . . . . . . . . . . . . . . . . . . . . . . 9811.6 Systematic Uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10111.7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .102

12 C O N C L U S I O N 107

B I B L I O G R A P H Y 109

1I N T R O D U C T I O N

The Large Hadron Collider (LHC) entered a new energy regime at Run 2 with proton–proton (pp) collisions at

ps = 13 TeV1. Events with multi-TeV jets showering in the

detectors, or tau-leptons and b-hadrons surviving passage through multiple active lay-ers of material, are now common place. These objects are also signatures for newphysics, including massive particles that decay to highly boosted bosons, whose ownsubsequent decay products are often reconstructed into one large-radius jet.

In the core of highly energetic hadronic jets, the average separation of charged par-ticles is comparable to the size of individual inner detector (ID) elements. This cancreate confusion within the algorithms reconstructing charged particle trajectories(tracks). Without careful consideration, the track reconstruction efficiency in thesedense environments will be limited, resulting in difficulties in identifying long-livedb-hadrons and hadronic tau lepton decays, or in the calibration of energy and jet massmeasurements. Mitigating such losses will increase the possibility of discovery of newphenomena and allow more detailed measurements of the newly opened kinematicregime. A dedicated optimization for dense environments was deployed in the ATLASreconstruction for the start of Run 2.

This thesis is structured in five major parts, where Part I introduces the StandardModel of Particle Physics, the LHC and the ATLAS experiment.

Part II gives a general overview of the track reconstruction algorithms describingthe performance of charged particle reconstruction in dense environments by the AT-LAS detector at the start of Run 2. Crucial components for dense environments arenoted. The quality of the expected performance is demonstrated in dedicated simu-lated samples of single particles and dijet events, and comparisons between simulationand data are performed in energetic jet events. Extending these Monte Carlo (MC)based studies, methods are introduced in Part III to probe the performance in data.One of these measures the inefficiency of the reconstruction in high pT jets, an envi-ronment of high charged particle multiplicity and collimation, by using the ionizationenergy loss (dE/dx) in the pixel detector.

Part IV explains how the mass of jets is reconstructed. It especially focuses on recentdevelopments extending the classical concept of a calorimeter-based mass to use infor-mation from the tracker - a concept which becomes feasible due to the improvementsin performance and uncertainties described in Part II. High jet mass resolution as wellas small related uncertainties are crucial for a plethora of physics analyses. Since themass can be used to identify jets originating from bosons, it is especially relevant forthe last part of the thesis.

Finally, a search for narrow resonances decaying hadronically into boosted bosonpairs is presented in Part V, exploiting the previously described developments in theATLAS track reconstruction as well as the novel jet mass definitions.

1 As common in Particle Physics, this thesis adopts “natural units” where the reduced Planck constant, thespeed of light in vacuum and the Boltzmann constant are set to unity (ħh = c = kB = 1) and the unit ofenergy is electron-volts (1 eV = 1.6×10−19 J).

1

Part I

T H E O RY & AT L A S E X P E R I M E N T

2T H E O RY

Throughout the history of science, individuals studying nature systematically havetried to describe it with mathematical models. The strength of such models lies notjust in describing the observed, but also in their ability to predict as yet unseen phe-nomena. In the study of fundamental particles, all objects and phenomena observedthus far can be described by the Standard Model of Particle Physics (SM). With anaccuracy that exceeds all current experimental limits, it is able to describe three of thefour fundamental interactions of nature: electromagnetism, the weak nuclear forceand the strong force. Using the formalism of Quantum Field Theory, the SM combinesaspects of quantum mechanics, classical field theory and Special Relativity. It fails toincorporate the fourth known force: gravity. The gravitational force is best describedby Einstein’s Theory of General Relativity [1]. Although it is the weakest of all funda-mental forces, it dominates on large scales, where it is able to describe the propertiesof time and space by linking gravitation to the geometry of the Universe. Unifyingall four forces into one theory is the greatest challenge for fundamental physics. Sec-tion 2.1 will briefly introduce the SM, while possible theories beyond it are saved forChapter V, where they are mentioned in the scope of a search for new physics. A morethorough description of the SM can be found in Reference [2].

2.1 T H E S TA N D A R D M O D E L O F PA RT I C L E P H Y S I C S

The SM introduces point-like particles with half-integer spin (fermions) whose inter-actions are mediated by integer spin particles (gauge bosons). These fermions includethree leptons with electric charge Q = −1e, the electron (e), muon (µ) and tau (τ),and corresponding neutral particles called neutrinos. They are complemented by theup, down, charm, strange, top and bottom quark with electric charge Q = + 2

3 e orQ = −1

3 e. For each of these particles, a corresponding anti-particle with invertedcharges exists. They are organized into three generations of up- and down-type quarks,and of neutral and charged leptons, with massive particles in each generation beingheavier than those in the previous generation. Each quark can exist in three differentstates of a quantum property called color charge, as first proposed by Gell-Mann [3].It is not known why the fundamental fermions are arranged into generations in thisway, nor whether additional generations of still heavier particles exist.

By unifying quantum electrodynamics (QED) and the theory of weak interactionsfirst proposed by Glashow, Weinberg and Salam [4–6], the SM is able to describe in-teractions through the electromagnetic and weak nuclear force in one, electroweak,theory. Quantum chromodynamics (QCD) is used to describe the strong force. One ofthe key concepts of all theories of the SM are local gauge symmetries. For example,the assumption that the electron field is invariant under local gauge transformationsleads to a new massless spin 1 gauge field, which materializes as the photon. Similarly,the gluons are introduced in the strong interaction and the weak bosons (W+, W−,

5

6 T H E O RY

Z) in the electroweak interaction. These gauge theories can be summarized with theU(1)×SU(2)×SU(3) group. The non-abelian nature of the SU(2) and SU(3) symme-try groups leads to the important effect of self-coupling of the respective bosons, forexample two oppositely charged W bosons can couple to a Z boson.

There is one major problem with this formalism: any mass term in the theory wouldviolate the chiral symmetry of the SU(2) group. By introducing the concept of spon-taneously symmetry breaking, Brout, Englert and Higgs [7, 8], were able to resolvethis issue by introducing an additional scalar field doublet. The so called Higgs mech-anism allows for massive leptons and weak bosons while retaining a massless photon.In addition, it predicts a massive excitation of the field, a Higgs boson, whose onlyfree parameter is its mass. Both the weak bosons and the Higgs boson were experi-mentally confirmed at CERN by the UA1 and UA2 experiments in 1983 [9, 10] and theATLAS and CMS experiments in 2012 [11, 12]. Figure 1 gives an overview of all theconstituents of the SM and their masses.

b

s

d

νe

νμνττ

μ

e

u

c

t

g

ZW

γ

H

Figure 1: Elementary particles of the Standard Model and their experimentally measuredmasses. The photon (γ), the gluon (g) and the neutrinos (ν) are assumed to bemassless. The outer ring consists of the fermionic particles, quarks in red and lep-tons in green. For each fermion a corresponding anti-particle exists. Gauge bosons,the mediators of the fundamental forces described by the Standard Model, are inthe blue ring. In the center is the Higgs boson, which gives rise to mass of the weakbosons.

3T H E L H C A N D T H E AT L A S E X P E R I M E N T

The ATLAS experiment [13] is a project with more than 20 years of history (the Letterof Intent was written in 1992). The detector itself is enormous, being 44 metres wide,22 metres in height and weighting 7000 tonnes. It is a multipurpose particle detector,designed to probe the forefront of Particle Physics, and is situated at the LHC [14]along with three other big experiments. CMS is another multipurpose detector andlike ATLAS it is able to measure, reconstruct and identify leptons, photons, quark jets,individual charged particles and primary and secondary vertices. ALICE and LHCbare designed for more specific purposes, which are recording events from heavy ioncollisions and detecting signatures from heavy flavor decays respectively. The maincomponents of the ATLAS detector are described in the following and can be seen inFigure 2.

Figure 2: A schematic drawing of the ATLAS detector. [13]

3.1 L A R G E H A D R O N C O L L I D E R AT C E R N

The LHC is a superconducting proton accelerator and collider installed in a 27 kilo-meters circumference underground tunnel at CERN. The 4-meter-wide tunnel wasbuilt for the Large Electron Positron (LEP) collider in 1985. The first studies for ahigh-energy pp collider in the LEP tunnel started in 1984. LEP2 was closed in 2000,and installation of LHC machine and experiments started in 2003. The first colli-sions at

ps = 900 (center-of-mass frame energy) GeV were recorded on November

23, 2009. The first collisions atp

s = 7 TeV took place on March 30, 2010, startinga long physics program. First collisions at

ps = 8 TeV were recorded on May 1st,

2012, and the machine performed spectacularly, delivering quickly increasing lumi-nosities [15]. On July 4th, 2012 the discovery of a Higgs-like boson was announced by

7

8 T H E L H C A N D T H E AT L A S E X P E R I M E N T

the ATLAS and CMS collaborations. In 2015 the collision energy was further increasedtop

s = 13 TeV and in 2016 the LHC surpassed its design luminosity reaching valuesof 1.37×1034 cm−2s−1.

3.2 C O O R D I N AT E S Y S T E M

The coordinate system and nomenclature used in the ATLAS experiment are brieflysummarized here, since they are used repeatedly throughout this thesis. The nominalinteraction point is defined as the origin of the coordinate system, while the beamdirection defines the z-axis and the x-y plane is transverse to the beam direction. Thepositive x-axis is defined as pointing from the interaction point to the center of the LHCring and the positive y-axis is defined as pointing upwards. The azimuthal angle φ ismeasured around the beam axis, and the polar angle θ is the angle from the beam axis.The pseudorapidity is defined as η = − ln (tan (θ/2)). In the case of massive objectssuch as jets, the rapidity

y = 0.5 ln [(E + pz)/(E − pz)] (1)

is sometimes used. The transverse momentum pT, the transverse energy ET, and themissing transverse energy Emiss

T are defined in the x-y plane unless stated otherwise.The distance ∆R in the pseudorapidity-azimuthal angle space is defined as ∆R =p

∆η2 +∆φ2 [13].

3.3 T H E M A G N E T S Y S T E M

The ATLAS magnet system is an arrangement of a central solenoid providing the mag-netic field for the ID, surrounded by three air-core toroid magnets which generate themagnetic field for the muon system. The generated field strength is 3.9 T and 4.1 Tfor the barrel and the end-cap toroids respectively. Each of the toroids consists of eightcoils which are aligned radially symmetric around the beampipe. This system, shownin Figure 3, has been one of the biggest engineering challenges of the ATLAS experi-ment due to its unusual layout and large size. The design is motivated by the goal tomeasure the muon momentum with an accuracy of 10% at 1 TeV.

The axial symmetric magnetic field of the solenoid has a strength of 2 T. It hasbeen designed to have as little density as mechanically possible, because any materialdegrades the performance of the calorimeters. This resulted in the housing of bothcalorimeter systems in the same vacuum vessel. The hadronic calorimeter and its sup-port structure also serve as the return yoke of the enclosed solenoid.

The resolution of particles reconstructed by the ID is directly dependent on thestrength of the magnetic field B in Tesla, since the reconstructed momentum p inunits of GeV is related to the radius of track curvature ρ in meters by

p = 0.299792458 · Bρ. (2)

As a consequence, the resolution of particles reconstructed in the ID increases athigher η due to the lower magnetic field provided by the solenoid in that phase space.

3.4 T H E I N N E R D E T E C T O R 9

end-cap toroid

barrel toroid

solenoid

tile calorimeter,return yoke

Figure 3: A schematic sketch of the ATLAS magnet system which shows all four magnets sys-tems as well as the tile calorimeter which acts as the return yoke of the solenoid. [13]

3.4 T H E I N N E R D E T E C T O R

The ATLAS ID [16] acts as a precise tracking and vertexing device. High precisionmeasurements with fine granularity sensors are required to make this possible withinthe high track density environment of LHC collisions. These are realized in the siliconpixel detector (including the recently added insertable B-layer or IBL) and silicon stripdetector (SCT), for which the detector elements closest to the beamline are only aboutthree centimeters from the interaction point. As a result of this setup each track crossesat least four pixel and four silicon strip layers. On average there are 37 measurementsper track, the majority of which are provided by the transition radiation tracker (TRT).The TRT enables tracks to be reconstructed out to large radii, although with a morelimited resolution. This limitation is countervailed by the large number of measure-ments as well as the large-radius, which still makes these measurements an importantaddition to the momentum measurement. The TRT is also used for particle identifica-tion and e/π separation with the help of the photons from transition radiation createdin its straw tubes.

The ID itself is 6.2 meter long and has a radius of 1.2 meter. It resembles the onion-like structure of classical collision experiments. In the end-caps the same componentsare present as described above (see Figure 4). The momentum resolution of recon-structed tracks is proportional to σ/pT ∼ 3.8×10−4pT(GeV)⊕0.015 [13].

3.4.1 The Pixel Detector

The pixel detector’s sensors are designed to provide a high-precision measurementas close as possible to the interaction point, in a pseudorapidity range of |η| < 2.5.For Run 2 of the LHC this capability was further enhanced by the addition of theIBL, which moves the first measurement from a radius of roughly 5 centimeters to 3centimeters. The former first pixel layer is still referred to as B-layer. The pixel detectors


Figure 4: A cut-out view of the ATLAS inner detector. [13]

measurements determine the possible resolution of the impact parameters, a quantitywhich is crucial for finding short lived particles such as b-mesons and τ-leptons. Thetwo dimensional readout has certain disadvantages. It requires the readout chips tobe of large area, with individual circuits for each pixel element. The proximity to theinteraction point comes at the cost of a very high radiation dose (around 158 kGy/year)compared to the other sub-detectors. The SCT only receives around 8 kGy/year. Thewhole system consists of 86 million detector elements with a size of 50 micrometers ×400 micrometers for the original pixel layers and 50 micrometers × 250 micrometersfor the IBL which provide an intrinsic resolution of 10 micrometers in R-φ and 100micrometers (50 micrometers for IBL) in z. A small number of pixels on each sensorare longer (500 micrometers for the IBL and 600 micrometers for the outer layers)to cover the gap between readout chips. The thickness of the original modules is 250micrometers and 200 micrometers for the IBL. As the number of readout channels permodule is not sufficient to read out every single pixel, as a compromise the last eightpixel rows are only connected to four readout channels which leads to an ambiguity in5% of the hits, which can be resolved offline. These pixels are called ganged. In additionto the four barrel layers at distances of 3.33, 5.05, 8.85 and 12.25 centimeters fromthe detector’s center, there are three disks on each side - at a distance of 49.5, 58.0and 65.0 centimeters from the detector’s center- completing the angular coverage. Alldetector elements have a readout such that the determination of the charge depositedby the traversing particles for each pixel by the so called Time-over-Threshold (ToT)method [17] is possible. All layers other than the IBL can do so with a 8-bit resolution.The IBL’s readout is only realized with a 4-bit resolution.

3.4.2 The Semiconductor Tracker

The SCT was designed to contribute to the measurement of momentum, impact pa-rameter and vertex position in the intermediate volume outside the pixel detector,

3.4 T H E I N N E R D E T E C T O R 11

Figure 5: Schematic view of a barrel pixel module. The main components are shown, includingthe MCC (module-control chip), the front-end (FE) chips, the Negative TemperatureCoefficient (NTC) thermistors, the high-voltage (HV) elements and the Type0 signalconnector. The side view on the bottom illustrates the bump-bonding of the siliconpixel sensors to the electronics. [13]

covering a pseudorapidity range of |η| < 2.5. Compared to the pixel, the still high-granularity is also crucial for good pattern recognition. In the barrel it consists of fourdouble-sided layers of silicon microstrip detectors at 29.9, 37.1, 44.3 and 51.4 cen-timeters from the detector’s center. This is complemented by nine disks at a distanceof 85.38, 93.40, 109.15, 129.99, 139.97, 177.14, 211.52, 250.50 and 272.02 centime-ters from the detector’s center. A small angle between the two sensors of each moduleenables measurements in the z coordinate. There is a total of 6.3 million sensors, eachmeasuring 80 micrometer × 12 centimeters, providing an intrinsic resolution of 16 mi-crometers in R-φ and 580 micrometers in z. This allows tracks separated by more thanapproximately 200 micrometers to be resolved. The readout consists of a front-end am-plifier and discriminator, followed by a binary pipeline. This binary output prohibitsmeasuring the deposited charge in any way.

3.4.3 The Transition Radiation Tracker

The TRT consists of straw tubes with a thickness of four millimeters and covers a rangein pseudorapidity up to |η| < 2.1. Only the R-φ coordinates are measured with an in-trinsic accuracy of 130 micrometers per straw. Still, it contributes significantly to themomentum measurement, since the low accuracy per measurement is compensatedby their abundance.. The 144 centimeters long straws are parallel to the beam axisin the barrel region, while in the end-cap region they are arranged radially in wheels.The total number of TRT readout channels is approximately 351 000. Electron identi-fication capabilities are added by measuring the signal of transition radiation photonscreated in a radiator between the straws. For this reason the straws are filled witheither a xenon or argon based gas mixture.

Baseboard TPG

Hybrid assembly

BeO facings (cooling side)Silicon sensors

BeO facings (far side)

Connector

Slotted washer

Datum washer

Hybrid assembly

BeO facings (far side)

Slotted washer

Silicon sensors

Baseboard TPG

Datum washer Connector

BeO facings (cooling side)

Figure 6: Drawing of a barrel SCT module, showing its components. [13]

3.5 T H E C A L O R I M E T E R

The ATLAS calorimeters (see Figure 7) are located outside of the solenoid. The AT-LAS electromagnetic calorimeter is based on Pb-LAr (lead-liquid argon) technology,with a novel accordion-like layout. It provides trigger, e/π identification and an en-ergy measurement with resolution of σ/E ∼ 10.1%/

pE ⊕ 0.17% [13] and coverage

in pseudorapidity |η| < 3.2 [15]. The forward liquid argon calorimeter provides cov-erage in the range 3.1 < |η| < 4.9 and a resolution of σ/E ∼ 29%/

pE ⊕ 0.04. The

hadronic central calorimeter is based on steel/scintillator tiles. It provides coverageup to |η|< 1.7 and a resolution of σ/E ∼ 50%/

pE⊕0.03. The end-cap calorimeters

use a copper/tungsten-LAr design, with coverage in pseudorapidity 1.5 < |η| < 3.2and resolution of σ/E ∼ 95%/

pE⊕0.08. The total mass of the ATLAS calorimeter is

about 4000 tons.

Figure 7: A cut-out view of ATLAS calorimeter detectors. [13]

3.6 T H E M U O N S P E C T R O M E T E R 13

3.6 T H E M U O N S P E C T R O M E T E R

The ATLAS muon spectrometer (see Figure 2) consists of gas-based muon chambersimmersed within a toroidal magnetic field originating from air-core coils. The muonsare detected with tracking chambers based on Cathode Strip (CSC) and MonitoredDrift Tubes (MDT) designs with a momentum resolution < 10% up to transverse mo-menta of about 1 TeV. Also triggering information is provided by its thin-gap (TGC)and resistive-plate (RPC) chambers.

3.7 T H E T R I G G E R S Y S T E M

The ATLAS trigger and data acquisition system is responsible for selecting for perma-nent storage and analysis∼ 1000 Events/s, out of the initial∼ 40 MHz. This two stagesystem is build up of the Level 1 Trigger (fast online electronics with decision times< 2.5 µs per event [15]) and the High Level Trigger (average decision time ∼ 40 ms)consisting of off-the-shelf computers and networking equipment.

3.8 T H E AT L A S C O M P U T I N G N E T W O R K

The ATLAS computing network is a distributed system which consists of the Tier-0at CERN where the data is recorded onto tape and the first pass reconstruction isperformed, 13 Tier-1 and 160 Tier-2 centers, which both participate in MC productionand support user analysis. For further details see Reference [18].

3.9 T H E AT L A S O F F L I N E S O F T WA R E - AT H E N A

The goal of the ATLAS offline software [19] is to process the multitude of measure-ments from the ATLAS detector, to deliver from this processed data the physics objectsto physicists, and to provide the necessary framework to analyze them in order toproduce physics results. Among the many additional requirements to the software,one shall be pointed out explicitly: the need to process the collected data within theavailable CPU and memory resources provided to ATLAS, while at the same time meet-ing the physics performance requirements. As the lifetime of the ATLAS experiment isdecades rather than years, its software needs to be able to adapt and evolve along withthe changing physics challenges. This motivates Athena, the common offline softwareframework in ATLAS. It adopts an object-oriented approach and is written in the C++language. Python is used to configure the run-time settings. The framework is builtprimarily from three types of software entities: algorithms, tools and services. Eachalgorithm is responsible for some basic task involving the input of data objects andthe output of new or modified objects. Tools and services are summoned by the algo-rithms as needed, and the same instances may be used by multiple algorithms. Athenaprocesses a given event by passing its constituent objects along a chain of algorithms,referred to as a sequence. Algorithms retrieve and store the data objects in a transientevent store called StoreGate. At any stage in the algorithm sequence this transient datacan be written to disk through a persistency service.

Part II

T R A C K R E C O N S T R U C T I O N I N D E N S E E N V I R O N M E N T S

4T R A C K R E C O N S T R U C T I O N

In order to understand the importance of the track reconstruction performance indense environments one has first to understand how charged particle reconstruction isperformed, which is explained in this chapter. Chapter 5 will then provide an overviewof specific improvements aimed at dense environments and Chapter 6 will demonstratethe resulting performance on MC simulated samples.

From the creation of the basic measurements in the ID known as clusters to the finalhigh resolution track fit, the reconstruction of charged particles’ trajectories, or tracks,is one of the most challenging computational and algorithmic tasks at the LHC. Afterthe creation of three-dimensional measurements, so called space-points, from clustersof pixels and SCT measurements, an iterative combinatorial track finder uses these asinput to create track candidates. An ambiguity solver is then used to process all trackcandidates and filter out those that are likely to be redundant. The outputs of thisalgorithm, in combination with measurements from the TRT, are the final ID tracksused for all further analyses. This chapter provides a detailed overview of the wholechain, while focusing on the most relevant parts of the algorithms for tracks in denseenvironments. A detailed description of aspects omitted here, such as how tracks areextended into the TRT, is given elsewhere [20].

4.1 P I X E L A N D S C T C L U S T E R I Z AT I O N

Single measurements on the same pixel or SCT module which share an edge or a cornerare grouped together by a connected component analysis (CCA) [21]. The geometricmean position of the group is used as the initial position estimate of the intersectionof the particle with the module. If the spatial separation of charged particles at themodule surface reaches the dimensions of a few pixels or strips, their charge depositsstart to overlap and the CCA will reconstruct a single merged cluster. Figure 8 illustratesthis behavior.

4.2 I T E R AT I V E C O M B I N AT O R I A L T R A C K F I N D I N G

In order to allow for as many combinations of space-points as possible, but still achievea rough momentum estimate, track seeds are built from three space-points. Seeds canbe built from space-points from either the pixel or SCT detector, or from combinationsof space-points from both detectors. This allows exactly four types of seeds with vary-ing purities. The purity is the fraction of seeds which in the end produce a good qualitytrack, and it is highest for seeds only formed from SCT space-points, followed by seedsformed only from pixel space-points. To maximize the efficiency of the track finding,seeds are processed ordered by their type. In addition, several selections are appliedon the seeds in order to further maximize their purity, including requirements on theminimum momentum of the seeds and their impact parameter. The latter are calcu-

17

18 T R A C K R E C O N S T R U C T I O N

(a) Isolated pixel clusters (b) Merged pixel cluster

Figure 8: Illustration of (a) isolated pixel clusters on a pixel module and (b) a merged pixelclusters due to very collimated tracks. Different colors represent charge deposits fromdifferent charged particles traversing the sensor and the particles trajectories areshown as arrows.

lated assuming a perfect helical trajectory of the particle through a uniform magneticfield. Finally, seeds passing these selections are also required to have one more space-point being compatible with their extrapolated trajectory. Space-points used by a seedpassing all these requirements are not considered further for other seeds, minimizingthe computational complexity of the algorithm. After seed creation, a combinatorialKalman filter [22] is used to create track candidates based on the seeds. Multiple trackcandidates can be built for each seed in case there are multiple possible extensions ofthe trajectory on a certain layer.

For a single muon about 13 combinations of space-points are created, but due to thequality requirements on the seeds, the Kalman filter is on average only called for 1.1of them. The resulting efficiency to reconstruct a muon track is higher than 99% [23].

4.3 T R A C K C A N D I D AT E S A N D A M B I G U I T Y S O LV I N G

Due to the combinatorial nature of the track finding, certain track candidates willhave clusters falsely assigned to them. Therefore the concept of ambiguity solvingis used: following the track candidate search, tracks are scored based on additionalquality criteria and afterward compared to each other. In case of ambiguities betweenmultiple track candidates the score decides which one to keep. Such approaches totrack finding have been around since the era of LEP experiments [24].

In the ATLAS offline reconstruction, track candidates are processed one after theother in order of descending score by such an ambiguity solver. The score needs tobe based on basic but robust measures of track quality, since the performance of theambiguity solver strongly depends on it. Initially, the logarithm of the track transversemomentum and the χ2 of the track fit are used as a base score of the track candidate.This gives preference to more energetic tracks and those with a good fit quality. Atthe same time, badly reconstructed tracks, which often have rather low transverse mo-mentum or a worse track fit are suppressed. In addition, every measurement assignedto a track candidate increases the score by a specific weight, based on the expected

4.3 T R A C K C A N D I D AT E S A N D A M B I G U I T Y S O LV I N G 19

number of measurements in the respective sub-detector and its resolution. As holes1

are an indication of a badly reconstructed track, they reduce the score.During the ambiguity solving, tracks are required to fulfill additional requirements.

One of them is limiting the number of pixel and SCT clusters which are used by mul-tiple tracks. These so called shared clusters can be an indication of clusters wronglyassigned to a track. At the same time, a merged cluster is expected to be assigned tomultiple tracks. To decide if multiple track candidates are allowed to use the samecluster, or if they should be penalized for doing so, one needs a method to identifymerged clusters. A first attempt to do so with the help of neural networks has been in-troduced to ATLAS during Run 1, and a detailed description of these neural networksis provided in Reference [25]. Correctly identifying merged clusters does improve thecluster assignment and track reconstruction efficiency, while at the same time decreas-ing the rate of badly measured tracks. Tracks may not share more than two clusters,and a cluster may not be shared by more than two tracks. Only track candidates whichwere accepted before the currently considered track are used to determine if a clusteris shared. In case an additional shared cluster would bring the current, or any of theprevious, track candidates above the maximum number of shared clusters, this clusteris removed from the track candidate. In such cases, the score of the track is recalcu-lated and it is added back to the list of track candidates which are to be processed. Thisscore would become zero, if the updated track candidate fails any of the basic mini-mum cluster or hole requirements. Track candidates with a score of zero are rejected.The flow of track candidates through the ambiguity solver is shown in a simplifiedform in Figure 9.

Fit tracks fulfilling minimum requirements

(Neural network used to predict cluster positions)

Calculate track scoreReject bad track score

Order tracks according to score (process from highest to lowest)

Input tracks

Create stripped-down track candidate

Accept track candidateor

Reject track candidate, due to too many shared clusters

(Neural network used to identify merged clusters)

too many holes too few clusters problematic pixel cluster(s)

Output tracks

Rejectedtracks

(if minimum requirements

fulfilled)

Figure 9:Sketch of the flow of tracks through the ambiguity solver.

Finally, a high resolution fit is performed for all tracks which were not rejected by theambiguity solver. This is only done at this point, to minimize the computing resourcesused by the fitter.

1 A hole is an intersection of the predicted particle’s trajectory with an active sensor element from whichno measurement is assigned to the track (inactive sensors are excluded from this definition).

20 T R A C K R E C O N S T R U C T I O N

4.4 T R U T H -B A S E D R E C O N S T R U C T I O N

An ideal detector level performance, without any inefficiencies due to the reconstruc-tion algorithms, is used as a benchmark for several results in the following sections.This is obtained from a truth-based2 reconstruction approach [26], which replaces allsteps described above apart from the final track fit. Clusters are assigned to track can-didates based on truth information skipping the pattern recognition completely. Theresulting track candidate must only fulfill a minimum silicon cluster requirement, beabove the same minimum pT as reconstructed tracks, and produce a successful trackfit. This results in perfect cluster assignment efficiency and in the maximum achiev-able reconstruction efficiency with the ATLAS detector. These tracks are referred to aspseudo-tracks.

4.5 T R U T H -B A S E D T R A C K Q UA L I T Y

In simulation, tracks are classified using a truth matching probability, defined as thefraction of measurements originating, from the same simulated particle. It is calculatedvia

PrParticle(Track) =

∑

wDet × nTrackMeas

∑

wDet × nPar t icleMeas

, (3)

where nTrackMeas is the number of measurements from the simulated particle used by the

track, nPar t icleMeas the actual number of measurements created in the detector by the sim-

ulated particle, and wDet a sub-detectors specific weight (10 for measurements in thepixel detector, 5 for the SCT and 1 for the TRT). A properly reconstructed track isrequired to have a truth matching probability above 0.5, and only those tracks areconsidered in the presented results from simulation. Tracks below that requirementare considered to be fake tracks.

2 Truth refers to all information taken directly from MC simulation, and which is not available in data.

5A L G O R I T H M I C I M P R O V E M E N T S T O T H E T R A C KR E C O N S T R U C T I O N

Chapter 4 described the ATLAS track reconstruction and touched upon some of theaspects crucial for tracking in dense environments (TIDE), for example the concept ofshared clusters. This setup was overhauled to prepare the offline reconstruction for thechallenges of LHC Run 2, and to improve the overall performance. Being the key stepfor dense environments, the ambiguity solver received the largest updates. Particularlythe handling of merged clusters was revised. This chapter gives an overview of thesechanges. Most of the content in this and the next chapter has been previously presentedin Reference [27] with significant contributions from the author of this thesis.

5.1 S A M P L E S

Simulated samples of single particle decays with very close-by charged decay productswere created to study the performance of the track reconstruction on event topologieswith highly collimated particles. This was achieved by generating parent particles witha flat transverse momentum spectrum from 10 GeV to 1 TeV within |η|< 1. A samplewith two highly collimated tracks (ρ → π+π−) and a sample consisting of a singleτ-lepton decaying to three charged hadrons (τ± → π+π−π±ντ) is used. In addition,a sample of single B+, decaying into a higher multiplicity of charged particles at asignificant radius from the luminous region, is used.

In the next Chapter, 6, additional samples are used of a Z ′ with a mass of 3 TeVdecaying to a pair of top quarks. These events were simulated using the PYTHIA 8 [28]generator with the AU2 [29] tuned parameters and the MSTW 2008 L0 [30] partondistribution functions (PDF) sets. Effects due to multiple pp interactions in the sameand adjacent bunch crossings (pile-up) are considered by overlaying, on average, 41minimum bias events simulated with PYTHIA 8. A value of 41 was the anticipatedaverage number for the pile-up per bunch crossing during Run 2, before the run began.

5.2 M E R G E D C L U S T E R S O N R E C O N S T R U C T E D T R A C K S

ATLAS uses an artificial neural network (NN) to identify merged pixel clusters duringthe reconstruction. ToT information in combination with shape and physical locationof clusters within the detector provides enough discriminating information for the NN.Due to the inherent randomness of the charge deposition process of charged particlesin the pixel detector, the ultimate performance of such a technique is limited. Emis-sion of low energy δ-rays can also be problematic, since they can lead to larger thanexpected ToT values and larger than average clusters. This can also have a negative im-pact on the spatial resolution achievable for pixel clusters. Failing to identify mergedclusters has a negative impact on the track reconstruction efficiency, can lead to in-

21

22 A L G O R I T H M I C I M P R O V E M E N T S T O T H E T R A C K R E C O N S T R U C T I O N

creased rates of shared clusters, and influence the number of holes on reconstructedtracks.

In the original setup used during Run 1, the NN’s output was used directly duringthe clusterization. Clusters identified as merged were physically copied (split). Basedon the information from additional NNs predicting the particles intersections withthe cluster, the positions of the copied clusters were updated. This approach has sev-eral weaknesses. During clusterization there is no available information on the trackcandidates associated with the cluster, so there exists no accurate information of theexpected incident angle of the particle with respect to the pixel sensors. Since thisinformation is highly correlated with the cluster shape and the measured ToT, it cansignificantly improve the NN performance. Figure 10 demonstrates this by comparingthe rate of incorrectly split one-particle clusters versus the rate of non-split two-particleclusters for NNs both with and without the track information. In the former case, thephysical position of the pixel module with respect to the luminous region position isused as an approximation of the average expected incident angle. For the same falsepositive rate for one-particle clusters, the efficiency to correctly identify merged clus-ters was 15% higher using the track information in Run 1.

Fraction of split 1-particle clusters

-310 -210 -110 1

Fra

ctio

n of

non

-spl

it 2-

part

icle

clu

ster

s

0

0.2

0.4

0.6

0.8

1

ATLAS Simulation

=7 TeVs

NN clustering

NN with track information

Figure 10: Fraction of pixel clusters falsely identified as merged (split) versus the fraction ofcorrectly identified merged clusters from two particles in simulation. The perfor-mance of the NN using the track information is compared to the NN only using theluminous region information. The star represents the working point for Run 1 andthe setup without track information. [25]

In addition to an underperforming NN, an additional load is put on the CPU andmemory resources due to the duplicating of clusters. In particular the resource ex-pensive combinatorial track finding suffers significantly under the additional clusters.Even for environments without real merged clusters, the NN has a non-negligible false-positive rate of around 10%, also visible in Figure 10. As a result, delaying the use ofthe NN until the ambiguity solver stage, where precise estimates of the incident anglesof the tracks are available, does not only improve its identification power, but simul-taneously reduces the number of track candidates that need to be processed by about10%. This results in a reduction of the required CPU and memory resources.

5.2 M E R G E D C L U S T E R S O N R E C O N S T R U C T E D T R A C K S 23

Not treating pixel clusters as isolated objects, but rather in a global picture includingtrack candidates, opens the door for additional developments. In the updated setup,the NNs output is only requested in case a cluster is assigned to more than one trackat the ambiguity solver stage. As such the false-positives have negligible impact onthe track reconstruction performance. Previously, clusters assigned to multiple trackswere only allowed as shared clusters, as described in Section 4.3. Now such clusterscan be used by multiple track candidates without penalty if the NN identifies them asmerged. This introduced the designation of shareable pixel cluster, while at the sametime retiring the previously used definition of split clusters. Clusters marked as sharedstill exist. These are not identified as shareable but used by multiple tracks, and aresubject to the restrictions mentioned in Section 4.3.

In cases where the NN fails to identify a merged cluster, due to its residual inefficien-cies, information from other clusters assigned to the track can be used to still mark itas shareable. Since the separation between particles usually increases with increasingradius from the interaction point, it is likely that a cluster merged in layer n is alsomerged in layer n− 1. If a cluster is shareable in layer n, and a cluster on the sametrack in layer n− 1 is assigned to multiple tracks, yet the NN does not identify it asmerged, it is still marked as shareable.

Motivated by these changes, a re-optimization of the minimum likelihood for a clus-ter to be identified as originating from two (three or more) particles based on the NNoutput is in order. Working points of 0.6 (0.2) maximize the track reconstruction effi-ciency while keeping the fake rate well under control. Further measures to maximizethe quality of the tracks passing the ambiguity solver are discussed in Section 5.3.

Figure 11 compares the distributions of three different categories of pixel clustercreated from the decay products of a single ρ and τ versus the pT of the mother parti-cles. The actual average number of merged pixel clusters is compared to the averagenumber of split clusters in the baseline and shared clusters in the TIDE reconstructionsetup. Based on truth information, the actual number of merged clusters is determinedby the number of clusters where more than one simulated particle deposited energyinto. With increasing pT, the separation between the two (three) charged decay prod-ucts of the ρ (τ) decreases, which is reflected by the increase of merged clusters la-beled as Ideal. Clusters marked as shareable in the TIDE setup closely follow the Idealdistribution, suggesting that this definition is adequate. As expected, at very low pT

both of them converge to zero. Contrary to this, the Baseline distribution suffers fromfalse-positives of the NN and stays well above zero.

At high pT, there is a lack of clusters marked as shareable/split in both reconstruc-tion setups, but much less prominent for the optimized TIDE setup. This effect is ex-acerbated in the τ decay, where merged clusters can be created by more than twoparticles.

Cluster assignment efficiency is an important measure of the performance of a trackreconstruction algorithm. It measures the efficiency with which a cluster created bya charged particle is correctly assigned to the reconstructed track associated to thatparticle. Figure 12 shows this efficiency for the IBL and B-layer. For the Baseline setupa sharp drop in efficiency is apparent for small minimum separations between parti-cles. For tracks from the ρ decay, the TIDE setup recovers 17% (13%) efficiency ata minimum separation on the order of a single pixel size for clusters on the B-layer

[GeV]T

pρ0 200 400 600 800 1000



0

0.5

1

1.5

2

2.5

3

3.5

4

4.5ATLAS Preliminary

-π+π→ρSimulation,

Ideal, mergedTIDE, shareableBaseline, split

(a) single ρ sample

[GeV]T

pτ0 200 400 600 800 1000



0

0.5

1

1.5

2

2.5

3

3.5

4


±π3τν→τSimulation,

Ideal, mergedTIDE, shareableBaseline, split

(b) single 3-prong τ sample

Figure 11: Average number of merged pixel clusters (based on truth information) and split(shareable) pixel clusters shown as a function of the (a) ρ and (b) τ transversemomentum.

(IBL). Compared to the resulting 98% cluster assignment efficiency for tracks from theρ decay, the overall efficiency for clusters in the B+ decay is lower due to the par-ticles non-zero lifetime and higher multiplicity in charged decay products. Still, theefficiency was increased at the smallest separation from 48% to 67% for clusters onthe IBL and from 62% to 94% for clusters on the B-layer. Similar improvements werealso achieved for the remaining layers of the pixel detector.

Figure 13 shows the average number of pixel clusters on track for events of single ρand 3-prong τ decays versus the pT of the parent particle. The average number of pixelclusters on tracks from the decay products of the ρ and 3-prong τ as a function of thepT of the parent particle are shown in Figure 13. Evidently, the number of clusters ontrack for the TIDE setup follows closely the expected number based on truth informa-tion, labeled Ideal, while for the Baseline setup the distribution diverges. This directlyrelates to the observations in Figure 11: since at high pT the Baseline setup fails toidentify most of the merged clusters, they are removed from the track candidate be-cause of the limitation on the maximum number of shared clusters. With large enoughboost, the 3-prong τ decay can occur after the IBL, so the expected average numberof clusters decreases with the pT of the τ. Both the Baseline and TIDE setup mirrorthat behavior, whereas the later again shows a significantly improved performance. Aresidual inefficiency in assigning all clusters to the reconstructed track remains for the3-prong decay, due to clusters created by more than two particles and the correspond-ing higher inefficiencies of the reconstruction. Compared to the Baseline the revisedTIDE reconstruction setup recovers on average up to 0.3 (0.45) pixel clusters on trackfrom the ρ (τ) decay.

5.2 M E R G E D C L U S T E R S O N R E C O N S T R U C T E D T R A C K S 25

Minimum Truth-Particle Separation at Layer 0 [mm]

-110 1

Clu

ster

Ass

ignm

ent E

ffici

ency

0.7

0.75

0.8

0.85

0.9

0.95

1

1.05

TIDELayer 0Layer 1

BaselineLayer 0Layer 1

ATLAS Preliminary-π+π→ρSimulation


Minimum Truth-Particle Separation at Layer 0 [mm]

-210 -110 1

Clu

ster

Ass

ignm

ent E

ffici

ency

0.4

0.5

0.6

0.7

0.8

0.9

1

TIDELayer 0Layer 1

BaselineLayer 0Layer 1

ATLAS Preliminary+Simulation B

(b) single B+ sample

Figure 12: Efficiency with which a cluster created by a charged decay product of the (a) ρ and(b) B+ is correctly assigned to the reconstructed track associated to that particle forthe IBL and B-layer as a function of the minimum truth particle separation at theIBL.

[GeV]T

pρ0 200 400 600 800 1000



2

2.5

3

3.5

4

4.5

ATLAS Preliminary-π+π→ρSimulation,

IdealTIDEBaseline


[GeV]T

pτ0 200 400 600 800 1000



2

2.5

3

3.5

4

4.5

ATLAS Preliminary±π3τν→τSimulation,

IdealTIDEBaseline


Figure 13: Comparison of the average number of pixel clusters on track for the Ideal, Baselineand TIDE reconstruction setup as function of the parent particle pT in single (a) ρand (b) τ events.

5.3 L I M I TAT I O N S O N S H A R E D C L U S T E R S

A first mention of how the updated ambiguity solver in combination with the NN canlimit the negative effects from shared pixel clusters was given in the previous section.The NN is not just able to identify merged clusters, but also to tag those created by asingle particle. Using this information, one can add a requirement based on the NNoutput, in order to allow a cluster to be shared between multiple tracks. Pixel clusterswith a predicted likelihood of being created by two particles of less than 0.05 cannotbe shared. This serves as an additional mean to limit the number of fake tracks. Also, atrack must have a minimum of nine clusters associated with it in order to allow any ofthese clusters to be shared with another track. This directly counteracts the fact thatmany mismeasured tracks arise from secondaries created in the outer layers of the ID,for example by inelastic interactions of the particles with the detector material. Tracksfrom these particles are likely to get clusters in the inner layers wrongly assigned tothem. Further, tracks are required to have a minimum pT of 1 GeV and at least fourSCT clusters associated with it which are not shared with another track in order tomark a pixel cluster assigned to it as shareable. Tuning was also performed on how theambiguity solver counts the number of shared clusters. In Run 1, once a track candidatepassed the solver, its shared cluster count was fixed. This could lead to cases wherea subsequent track increases the number of shared clusters on an already acceptedtrack, and at a certain point, the already accepted track could violate the maximumshared cluster requirement. In the optimized setup, no track will be accepted if it eitherfails the maximum shared cluster requirement itself, or causes any previously acceptedtrack to do so, which eventually favors higher scored tracks.

In summary, major changes compared to Run 1 are:

1. Delaying the evaluation of the pixel cluster NN from the clusterization to theambiguity solving.

5.3 L I M I TAT I O N S O N S H A R E D C L U S T E R S 27

2. Introducing the denomination of shareable cluster, while at the same time abol-ishing split clusters:

• Pixel clusters identified by the NN as merged, and assigned to multipletracks, are shareable.

• Pixel clusters assigned to multiple tracks but not identified by the NN asmerged are shareable if the next outward cluster of the track is a shareablecluster.

• Shareable clusters are not penalized for being used by multiple tracks (splitclusters were not allowed to do so at all).

3. Existing cuts tuned to optimize efficiency and fake rejection:

• Minimum NN likelihood of two particle hypothesis for shareable cluster:0.6

• Minimum NN likelihood of more than two particle hypothesis for shareablecluster: 0.2

• Maximum NN likelihood of two particle hypothesis for shared cluster: 0.05

• Minimum number of pixel plus SCT clusters required to allow a cluster tobe shared by a track: 9

4. Additional cuts to allow designation of a cluster as shareable (suppresses faketracks):

• Minimum number of SCT clusters used only by the track candidate: 4

• Minimum track candidate pT: 1 GeV

Due to the lack of charge information from the detector, the track reconstructiondoes not posses the possibility to identify merged SCT clusters in the same way as itdoes for pixel clusters. Since loosening the shared cluster requirement (maximum twoper track) would disproportionately increase the rate of badly reconstructed tracks,this puts an inherent limit on the currently achievable track reconstruction efficiencyin dense environments. Assuming perfect identification of merged pixel clusters, Fig-ure 14 demonstrates this effect based on the efficiency of reconstructing all tracks ofthe charged decay products from ρ and τ decays. This result is not based on actual re-constructed tracks, but rather on the truth-based tracks introduced in Section 4.4. Theonly factor limiting the efficiency is the requirement of a minimum number of clusterson each track, and the varying requirement on the maximum number of shared SCTclusters. At higher momentum the decay products of theρ and τ become so collimated,that on average even at the radius of the SCT their clusters merge.

Disentangling this effect, Figure 15 shows the actual efficiency to reconstruct all pri-mary tracks where none are expected to share more than two clusters in the SCT. Thecomparison between the Baseline and TIDE setup at high pT reveals a clearly higherefficiency of the later, due to the improvements in handling merged pixel clusters. An-other factor limiting the achievable efficiency are inelastic interactions of the particleswith the detector material. Therefore, the algorithmic efficiency without these inter-actions is almost 100% for the ρ, and slightly lower for the τ. A residual inefficiencyremains at high pT, where errors of the pattern recognition and the NN can lead to


[GeV]T

pρ0 200 400 600 800 1000

Rec

onst

ruct

ion

Effi

cien

cyρ

Idea

l

0.2

0.4

0.6

0.8

1

1.2 ATLAS Preliminary-π+π→ρSimulation,

All 8 Shared SCT clusters≤ 6 Shared SCT clusters≤ 4 Shared SCT clusters≤ 2 Shared SCT clusters≤ 0 Shared SCT clusters≤


[GeV]T

pτ0 200 400 600 800 1000

Rec

onst

ruct

ion

Effi

cien

cyτ

Idea

l

0.2

0.4

0.6

0.8

1

1.2 ATLAS Preliminary±π3τν→τSimulation,

All 8 Shared SCT clusters≤ 6 Shared SCT clusters≤ 4 Shared SCT clusters≤ 2 Shared SCT clusters≤ 0 Shared SCT clusters≤


Figure 14: Truth-based efficiency to reconstruct all tracks of the charged decay products from(a) ρ and (b) τ decays. It is assumed that merged pixel clusters will always beidentified as such, and only the requirement on the number of shared SCT clustersis varied.

incorrect cluster assignments, which ultimately lead to a failure to classify the trackas being good. In events where secondary particles from interactions add additionalclusters, and possible confusion, this effect is more evident.

5.3 L I M I TAT I O N S O N S H A R E D C L U S T E R S 29

[GeV]T

pρ0 200 400 600 800 1000

Alg

orith

mic

Effi

cien

cyρ

0.9

0.95

1


-π+π→ρSimulation, 2 Shared SCT Clusters≤

No Secondaries

BaselineTIDE

[GeV]T

pρ0 200 400 600 800 1000

Alg

orith

mic

Effi

cien

cyρ

0.9

0.95

1


-π+π→ρSimulation, 2 Shared SCT Clusters≤

BaselineTIDE


[GeV]T

pτ0 200 400 600 800 1000

Alg

orith

mic

Effi

cien

cyτ

0.8

0.9

1


±π3τν→τSimulation, 2 Shared SCT Clusters≤

No Secondaries

BaselineTIDE

[GeV]T

pτ0 200 400 600 800 1000

Alg

orith

mic

Effi

cien

cyτ

0.8

0.9

1


±π3τν→τSimulation, 2 Shared SCT Clusters≤

BaselineTIDE


Figure 15: Efficiency to reconstruct all primary tracks from (a) ρ or (b) 3-prong τ decaysas a function of the parent pT. Based on truth information, none of the tracks areexpected to share more than two SCT clusters. Left: primary particles generate nosecondary particles through inelastic interactions with the detector material. Right:no requirement on the generation of secondary particles.

6R U N 2 P E R F O R M A N C E A N D S T U D I E S W I T H T H E I T K L AY O U T

6.1 J E T R E C O N S T R U C T I O N

Jets are collimated streams of particles resulting from the production and hadroniza-tion of high energy quarks and gluons. Energy deposits from these particles in thecalorimeter cells, or the reconstructed charged particle tracks, serve as input to the jetreconstruction. Most calorimeter-based methods reconstruct the jet four-momentumfrom the magnitude, direction and topology of the deposited energy. ATLAS uses topo-logically adjacent calorimeter-cell clusters (topo-clusters [31]) as inputs to the jet find-ing algorithms, and treats them as massless four-vectors of energy E =

∑

Ecel l . The jetfinding then tries to combine four-vectors which are likely to originate from the samequark or gluon using two pT weighted distances between them. The first distance di j

between two four-vector i and j can be defined as

di j = min(paTi , pa

T j)×∆Ri j

R, (4)

where a is an algorithm specific coefficient, ∆Ri j is the angular distance betweenthe two four-vectors and R is the radius parameter defining the final size of the recon-structed jet. The second distance is defined as diB = pa

Ti and describes the separationbetween a four-vector and the beam axis in the momentum space.

Each jet algorithm based on these distances finds the minimum of di j and diB. Ifdi j is the minimum, the two four-vectors i and j are combined into one, and removedfrom the list of four-vectors. If diB is the minimum, i is removed from the list and isdesignated as being a final jet. This is repeated until all four-vectors, within ∆Ri j < R,are assigned to a final jet. Two different values of a are used within the context ofthis thesis: a = 2 and a = −2. They correspond to the so-called kt and anti-kt algo-rithm [32]. The former prefers to assign lower pT particles to the jet first, which makesit sensitive to noise effects as pile-up. However, it does perform well when resolvingsubjets (i.e. jets reconstructed with a smaller R parameter within a jet reconstructedusing a larger R parameter). The negative a value of the anti-kt algorithm results in re-constructed jets most sensitive to high pT contributions, which makes it stable againstany noise effects. At the same time, it is unable to resolve any substructure [33] bybuilding small R jets. Figure 16 shows the same truth-level particles clustered by thetwo different algorithms. While anti-kt has a very robust reconstructed jet area, thektalgorithms jet’s area clearly varies significantly.

Small-radius jets are built from topo-clusters, at the electromagnetic scale and thencorrected on average for the effects of pile-up, using the FASTJET [34] implementationof the anti-kt algorithm with a distance parameter R = 0.4. Next, a calibration to thehadronic scale is applied as documented in Reference [35]. To account for differencesin the response determined by in-situ methods [35], a final correction is applied todata.

31

32 R U N 2 P E R F O R M A N C E A N D S T U D I E S W I T H T H E I T K L AY O U T

(a) ktalgorithm (b) anti-kt algorithm

Figure 16: A sample event with topo-clusters in the φ-y plane clustered with two different jetalgorithms, illustrating in color the areas of the resulting final jets. [32]

Large-radius jets are also formed with the anti-kt algorithm, but with a distanceparameter R = 1.0. The local cell signal weighting (LCW) method [31] is used tocalibrate calorimeter-cell clusters to the hadronic scale, which are then used as input toform large-radius calorimeter jets. To reduce noise effects, like pile-up, the kt algorithmis used to built subjets with a distance parameter Rsub = 0.2 within the large-radiusjet, and removing any constituents of subjets with pT less than 5% of the large-radiusjet pT. This procedure is referred to as trimming [36]. Simulation based energy andη dependent calibration factors are used to further calibrate energy and η of the jetand to remove residual detector effects [35, 37, 38]. Finally, the large-radius jet massis calibrated as described in Section 10.2. For all results in this chapter, large-radiuscalorimeter jets are required to have pT > 200 GeV with |η|< 2.0. Simulated particleswith a large enough lifetime (cτ > 10 mm), excluding muons and neutrinos, are usedas inputs for truth jets which are built from the generated truth particles.

If in simulation a truth W /Z-boson (top quark) is associated to the untrimmed areaof the large-radius calorimeter jet using the ghost association method [39], the jet islabeled a W /Z-jet (top-jet).

6.2 T R A C K R E C O N S T R U C T I O N P E R F O R M A N C E I N T H E C O R E O F J E T S

In the previous sections, the differences in performance between the Baseline andTIDE setup have been shown for single particle decays. These samples demonstratedthe performance in very specific topologies, thus they are not fully representative ofthe expected performance on physics data. The most abundant objects for which thetrack reconstruction encounters dense environments are high momentum jets. Fig-ure 17 (a) compares the average number of IBL clusters on tracks inside jets from aZ ′→ t t̄ decay (introduced in Section 5.1). These jets are required to have pT greaterthan 100 GeV and to be within the acceptance of the ID (|η| < 2.5). For small sep-arations between the track and the jet axis, both reconstruction setups show a dropin the average number of clusters on track, while the effect is less prominent for theTIDE setup. Figure 17 (b) compares the distributions of clusters previously defined assplit with the distribution of shareable clusters in the TIDE setup. While the Baseline

6.2 T R A C K R E C O N S T R U C T I O N P E R F O R M A N C E I N T H E C O R E O F J E T S 33

distribution does not exhibit any physical trend, the TIDE distribution shows a slightincrease in the number of shareable clusters with decreasing ∆R(jet,track), followedby a sharp rise for tracks in the very core of the jet. This rise is correlated with theincreased particle density in the core.

R(jet,trk)∆

0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2

<In

nerm

ost L

ayer

Clu

ster

s>

0.8

0.9

1

1.1

1.2

1.3

1.4


=13 TeV, Z’(3 TeV)sSimulation, TIDEBaseline

(a)

R(jet,trk)∆

0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2

<S

plit/

Sha

reab

le P

ixel

Clu

ster

s>

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8 ATLAS Preliminary=13 TeV, Z’(3 TeV)sSimulation,

TIDEBaseline

(b)

Figure 17: (a) Average number of clusters from the innermost pixel layer assigned to primarytracks and (b) average number of split/shareable pixel clusters assigned to primarytracks as a function of the angular distance between the track and the jet axis. Itis required that the production vertex of the particle is before the innermost pixellayer and that the jet pT > 100 GeV.

A comparison of the Baseline and TIDE reconstruction efficiency for charged primaryparticles is presented in Figure 18 as a function of jet pT. The Baseline setup showsa significant drop in efficiency at higher jet pT, while the TIDE setup is much morerobust.

In jet cores, that is, for angular distances between the charged particle and thejet axis of about 0.05 and less, a residual reconstruction inefficiency is likely. Clus-ters created by multiple particles are common there, similar to tracks from the pre-viously discussed B+ and τ decays. The track reconstruction efficiency for charged


[GeV]T

Truth Jet p

100 200 300 400 500 600 700 800 900 1000

Tra

ck R

econ

stru

ctio

n E

ffici

ency

0.78

0.8

0.82

0.84

0.86

0.88

0.9

0.92

0.94

0.96 ATLAS Preliminary=13 TeV, Z’(3 TeV)sSimulation,

TIDEBaseline

Figure 18: Track reconstruction efficiency of primary particles with a production vertex beforethe IBL dependent on the jet pT.

primary particles as a function of ∆R(jet,track) is shown in Figure 19 (a) for jets withpT 450–750 GeV. Charged particles used in this result need to have a minimum pT

of 2 GeV, must have been created within a radius of 100 mm from the center of thedetector, and must not decay or undergo inelastic scattering within the SCT volume(Rpart

decay > 600 mm). Due to displaced decays of heavy flavor quarks, the observed inef-ficiency at low angular separations is worse for b-jets. An improvement of 10% (14%)is achieved in the core of light (b) jets with the TIDE reconstruction setup. With theTIDE setup, the same track reconstruction efficiency within b jets is achieved than waspreviously possible in the Baseline setup for light jets. Figure 19 (b) shows, with thesame selections as Figure 19 (a), the track reconstruction efficiency for charged pri-mary particles as a function of the production radius of the particle. Two main effectswill reduce the efficiency at higher production radii. Firstly, decays occurring immedi-ately before a pixel layer will lead to an increased probability of merged clusters fromseveral charged particles. Secondly, the total number of created clusters in the detectorwill decrease if the particle is created after the nth pixel layer. The latter results in asharp drop in efficiency down to around 80% and 50% at a production radius above33 and 50 mm (the radii of the IBL and B-layer). For particles created close to the IBL,an increase in efficiency of 17% is achieved with the TIDE reconstruction setup.

6.3 I M P L I C AT I O N S F O R F L AV O R TA G G I N G

Several techniques exist to identify jets containing b-hadrons, though all of them insome way seek the characteristic displaced production vertex. This is usually achievedthrough an analysis of the impact parameters of the tracks or a direct reconstructionof the displaced vertex [40]. ATLAS combines the results from multiple taggers, eachusing a different approach to flavor tagging, through a multivariate approach to obtainthe final jet flavor variable. The impact parameter based taggers are especially sensitiveto mis-assignments or the lack of clusters on the reconstructed tracks, as both candirectly impact the impact parameter resolution. An improved performance due to theimprovements to the track reconstruction is expected. To some level, secondary vertex

6.3 I M P L I C AT I O N S F O R F L AV O R TA G G I N G 35

R(jet,trk)∆

-210 -110

Tra

ck R

econ

stru

ctio

n E

ffici

ency

0.7

0.75

0.8

0.85

0.9

0.95

1

ATLAS Preliminary=13 TeV, Z’(3 TeV)sSimulation,

>2 GeVpart

T<750 GeV, p

jet

T450600 mmpart

decay<100 mm, R

partprodR

TIDELight-Jets

B-Jets

Baseline Light-Jets

B-Jets

(a)

Truth Particle Production Radius [mm]1 10 210

Tra

ck R

econ

stru

ctio

n E

ffici

ency

0.5

0.6

0.7

0.8

0.9

1


>2 GeVpart

T<750 GeV, p

jet

T450600 mmpart

decay R(jet, part)<0.2, R∆TIDEBaseline

(b)

Figure 19: Track reconstruction efficiency for charged primary particles as a function of (a) theangular distance between the particle and the axis of the jet and (b) the productionradius of the particle.


based taggers will also see an improvement, since discriminating variables derivedfrom the associated tracks are used.

In order to evaluate the actual possible gains for flavor tagging due to the improve-ments of the track reconstruction, one would need to re-optimize and tune the wholeset of used taggers because of the change in the inputs. Since this is beyond the scopeof this thesis, only the change in performance of the IP3D tagger [41], which is ex-pected to be most sensitive to the changed track inputs, will be presented. No opti-mization was performed on it, and the results shown in the following therefore serveas an lower limit on the actual resulting improvement. The b-jet tagging-efficiency ,Figure 20 (a), and the light-jet rejection at the 70% efficiency working point, Figure 20(b), are shown for different truth jet pT for jets from the Z ′ decay. An increased effi-ciency, especially at high jet pT, with about the same light-jet rejection is observed forthe TIDE setup compared to the Baseline. Figure 21 shows the b-jet efficiency versusthe light-jet rejection for all jets with a pT > 100 GeV. For a light jet rejection equalto that at the 50% and 80% b-jet identification efficiency working points of the IP3Dtagger, a relative increase in efficiency of 13% and 7% is obtained when using the TIDEtrack reconstruction setup.

6.4 P E R F O R M A N C E O F T R A C K R E C O N S T R U C T I O N I N D E N S E E N V I R O N M E N T S

W I T H T H E I T K

The LHC will undergo a significant upgrade with the goal of increasing its integratedluminosity by a factor of 10 in the years after Run 3 [42]. To achieve this, an eventpile-up of up to 200 inelastic pp interactions per bunch crossing is necessary. To pre-pare the ATLAS detector for this harsh environment, the entire ID will be replaced bythe new Inner Tracker (ITk) [43]. One of the main tasks of the ITk will be the recon-struction of tracks in jets in the currently uncovered very forward region of η up to 4.0.This will be crucial if ATLAS is to be able to reject pile-up jets in that region by meansof track-to-vertex matching [43]. The expected tracking performance in dense envi-ronments is one of the main areas of study for the ITk preparations. Figure 22 showsthe efficiency to reconstruct all tracks from the charged decay products of a 3-prongτ-decay as a function of the τ pT for a specific candidate tracker layout. This layoutdiffers significantly from the current ATLAS ID. Most prominently it is a silicon-onlytracker made out of five pixel layers and four strip layers with strip doublet modules inthe barrel, with a novel inclined sensor arrangement in the forward region. Figure 22shows this candidate layout and its details and the reconstruction setup are describedin Reference [44]. A clear degradation of the performance is visible in the very forwardregion η > 2.7. Such studies are crucial in determining the capabilities of proposedfuture detector layouts.

6.5 C O N C L U S I O N

Optimizations provide a vastly improved track reconstruction setup in terms of perfor-mance in dense environments for Run 2. An improved performance is demonstratedfor highly collimated tracks from decays of single particles, as well as in the more phys-ical environment of hadronic Z ′ decays in the presence of event pile-up. Up to 10%

6.5 C O N C L U S I O N 37

[GeV]T

Truth Jet p

0 200 400 600 800 1000 1200 1400

B-J

et E

ffici

ency

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1ATLAS Preliminary

=13 TeV, Z’(3 TeV)sSimulation, IP3D

TIDEBaseline

(a)

[GeV]T

Truth Jet p

0 200 400 600 800 1000 1200 1400

Ligh

t-Je

t Rej

ectio

n

1

10

210 ATLAS Preliminary=13 TeV, Z’(3 TeV)sSimulation,

IP3D

TIDEBaseline

(b)

Figure 20: (a) b-jet efficiency and (b) light-jet rejection using the IP3D tagger at the 70%working point with input tracks from the Baseline and TIDE reconstruction setupsas a function of truth-jet pT for jets within |η|< 2.5.

B-Jet Efficiency

0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Ligh

t-Je

t Rej

ectio

n

1

10

210


IP3D

TIDEBaseline

Figure 21: Light-jet rejection as a function of b-jet identification efficiency using the IP3Dalgorithm with input tracks from the Baseline and TIDE reconstruction setups. Jetsare required to have pT greater than 100 GeV and |η|< 2.5.

[GeV]T

Truth Tau p

0 200 400 600 800 1000 1200 1400 1600 1800 2000

3-pr

ong

Rec

onst

ruct

ion

Effi

cien

cy

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1.1

1.2

ATLAS Simulation Preliminary=0µ, ±π3τν->τSingle

ITk Inclined| < 1.0τη|

| < 2.7τη1.0 < || < 4.0τη2.7 < |

Figure 22: Efficiency to reconstruct all primary tracks from 3-prong τ decays as a function ofthe parent pT with a proposed ITk tracker layout for three different η regions.

6.5 C O N C L U S I O N 39

Figure 23: Diagram showing simulated energy deposits in active layers for a candidate ITklayout with an “inclined" pixel barrel, shown in the R-z plane. The pixel tracker isin red, while the strip tracker is blue. [44]

more pixel clusters are associated to tracks in the core of high pT jets, which results ina much more robust track reconstruction efficiency. This improvement becomes mostapparent for charged particles with a high production radius (>30 mm), where 17%efficiency is recovered, as well as in the core of high pT b- and light-jets, where 10%and 14% efficiency is recovered. The higher quality and efficiency of reconstructedtracks directly boosts the performance of several derived physics objects, which wasdemonstrated with the example of flavor tagging, and is also relevant for the jet massdescribed in Chapter 10. For flavor tagging a 7–13% improvement in b-jet efficiency isachieved for a fixed light-jet rejection for jets with pT > 100 GeV, using an IP3D taggerwhich has not been re-optimized.

Part III

M E A S U R I N G T R A C K R E C O N S T R U C T I O N P E R F O R M A N C E I ND ATA

7P R O P E RT I E S O F R E C O N S T R U C T E D T R A C K S I N D E N S EE N V I R O N M E N T S I N R U N 2 D ATA

The previous part of this thesis described the track reconstruction in dense environ-ments and discussed its performance based on MC simulated samples. Due to imper-fect modeling of the multiplicity and angular distribution of charged particles withina jet in simulation, certain discrepancies with data are expected. Simulation is alsoknown to describe the properties of pixel and SCT clusters observed in data only toa certain level. This chapter contains comparisons of basic properties of tracks in thecore of high pT jets between data and MC simulation to probe the overall agreementbetween the two. Chapter 8 describes the use of data to evaluate the performanceof the NN identifying merged clusters and Chapter 9 discusses a measurement of theresidual inefficiency of the track reconstruction in data.

The comparisons in this chapter are mainly performed as a function of∆R(jet,track),the separation between the track and the jet axis. Tracks are associated to the calorime-ter jets following the ghost association procedure. Where results from data are pre-sented, 3.2 fb−1 of 2015 LHC collision data is used and standard data quality require-ments are applied to select the luminosity blocks where all sub-systems of the detectorwere fully operational. A simulated MC sample generated with the PYTHIA 8 with theA14 [29] tuned parameters and the NNPDF23LO [45] PDF sets is used. Interactionsof all generated particles with the detector are fully simulated through GEANT 4 [46]and simulation is reweighted to make sure that the average number of interactionsper bunch crossing and the jet η spectra match those in data.

Besides considering the average number of pixel and IBL clusters on track as a func-tion of the distance of the track to the jet axis, the main interest lies in the averagenumber of shared and shareable clusters. These numbers are strongly depended on thecharged particle density in the jet. The average number of tracks versus the angulardistance from the jet axis in data and MC is compared in Figure 24. In data, about 5to 10% less tracks are observed as in simulation and the discrepancy between the twobecomes larger with increasing jet pT.

Figure 25 shows the average number of shared pixel and SCT clusters on tracks injets for data and MC versus the angular distance from the jet axis. For small distancesa sharp increase in the number of shared pixel clusters is visible. Although the SCTsensors are located at much higher radii than the pixel sensors, a large numbers ofshared clusters exist. This is due to the larger sensor dimensions of the SCT strips, aswell the fact that no NN can be utilized to identify merged SCT clusters, as discussedfor the pixel clusters in Section 4.3. The MC simulation on overall agrees well withdata in the individual bins of jet pT, but the average number of shared pixel clustersdeviates from it with decreasing angular separation between the track and the jetaxis. A maximum discrepancy of around 15% is observed for ∆R(jet,track) < 0.04.Considering the agreement observed in Figure 24 and the fact that a simplified chargedeposition model is used in simulation, such differences for pixel clusters are expected.

43

44 P R O P E RT I E S O F R E C O N S T R U C T E D T R A C K S I N D E N S E E N V I R O N M E N T S I N R U N 2 D ATA

R(jet,trk)∆

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4<

Num

ber o

f Tra

cks>

1

2

3

4

5

6<400 GeVjet

Tp≤200

<800 GeVjet

Tp≤600

<1400 GeVjet

Tp≤1200

<2000 GeVjet

Tp≤1800

jet

s=13 TeV, 3.2 fb-1

|η |<1.2

Data MC

R(jet,trk)∆0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Dat

a/M

C

0.80.9

11.11.2

Figure 24: The average number of tracks in different bins of angular distance from the jet axis.Data and MC are compared in bins of jet pT.

Figure 26 shows the average number of IBL clusters, and the fraction of pixel clusterswhich are identified as merged, on tracks in jets for data and MC versus the angulardistance from the jet axis. For small separations the number of clusters shows a slightdrop, while the number of shareable clusters steeply rises. MC and data show goodagreement in the individual bins of jet pT.

P R O P E RT I E S O F R E C O N S T R U C T E D T R A C K S I N D E N S E E N V I R O N M E N T S I N R U N 2 D ATA 45

R(jet,trk)∆

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4



0.02

0.04

0.06

0.08

0.1

0.12 <400 GeVjet

Tp≤200

<800 GeVjet

Tp≤600

<1400 GeVjet

Tp≤1200

<2000 GeVjet

Tp≤1800

jet

s=13 TeV, 3.2 fb-1

|η |<1.2

Data MC

R(jet,trk)∆0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Dat

a/M

C

0.80.9

11.11.2

(a)

R(jet,trk)∆

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

<S

CT

Sha

red

Clu

ster

s>

0.2

0.4

0.6

0.8

1

1.2<400 GeVjet

Tp≤200

<800 GeVjet

Tp≤600

<1400 GeVjet

Tp≤1200

<2000 GeVjet

Tp≤1800

jet

s=13 TeV, 3.2 fb-1

|η |<1.2

Data MC

R(jet,trk)∆0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Dat

a/M

C

0.80.9

11.11.2

(b)

Figure 25: The average number of shared (a) pixel and (b) SCT clusters for primary tracks(with a production vertex before the first layer) as a function of the angular distancebetween the track and the jet axis. Data and MC are compared in bins of jet pT andshow an overall excellent agreement.

46 P R O P E RT I E S O F R E C O N S T R U C T E D T R A C K S I N D E N S E E N V I R O N M E N T S I N R U N 2 D ATA

R(jet,trk)∆

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

<In

nerm

ost P

ixel

Lay

er C

lust

ers>

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6 <400 GeVjet

Tp≤200

<800 GeVjet

Tp≤600

<1400 GeVjet

Tp≤1200

<2000 GeVjet

Tp≤1800

jet

s=13 TeV, 3.2 fb-1

|η |<1.2

Data MC

R(jet,trk)∆0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Dat

a/M

C

0.80.9

11.11.2

(a)

R(jet,trk)∆

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4



0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

2.2<400 GeVjet

Tp≤200

<800 GeVjet

Tp≤600

<1400 GeVjet

Tp≤1200

<2000 GeVjet

Tp≤1800

jet

s=13 TeV, 3.2 fb-1

|η |<1.2

Data MC

R(jet,trk)∆0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Dat

a/M

C

0.80.9

11.11.2

(b)

Figure 26: The average number of (a) innermost pixel layer clusters and (b) number of share-able pixel clusters on primary tracks (with a production vertex before the first layer)as a function of the angular distance between the track and the jet axis. Data andMC are compared in bins of jet pT and show an overall excellent agreement.

8M E A S U R I N G T H E P E R F O R M A N C E O F T H E P I X E L C L U S T E RN E U R A L N E T W O R K I N D ATA

It was shown in Chapters 4 and 5 that for tracks in dense environments a good perfor-mance and understanding of the NN pixel clustering used in the ATLAS track recon-struction is vital. The basic performance of this method has been presented previouslybased on MC simulation [25]. However, no study of the ability of the reconstructionsoftware to identify merged clusters in data has been performed. Such a measurementfrom data is crucial to obtain confidence in the agreement between the output of theNN in data and MC simulation, and the related uncertainties. Two independent tech-niques are introduced in Section 8.1 which allow such a measurement, and resultsare presented in Section 8.2. The study is performed on a subset (43.4 pb−1) of thedata set used in Chapter 7 and the same dijet MC simulation is used. Results from thischapter have been presented in Reference [47] with significant contributions from theauthor of this thesis.

8.1 M E T H O D S

The principal of both methods measuring the ability of the ATLAS track reconstruction,and its NN pixel clustering, to identify merged clusters is simple: pairs of collimatedtracks are built, of which each has to pass the following selection:

• pT > 1 GeV,

• |η| < 2.5,

• Longitudinal impact parameter, |z0|, with respect to the primary vertex1 < 5mm,

• Number of clusters in the pixel and SCT ≥ 7,

• Number of holes in the pixel and SCT ≤ 2,

• Number of holes in the pixel ≤ 1,

• Number of shared clusters in the pixel and SCT ≤ 1,

To ensure that only highly collimated pairs of tracks are considered, each of themmust be within a radial separation, ∆R(track,track), of 0.1. Every pair in every eventsatisfying the above criteria is considered for this study, which results in three mil-lion pairs of tracks from data. The simulation is normalized to this number. Figure 27compares the leading and sub-leading track pT distribution between data and simula-tion. Like for other kinematic variables of the track pairs, such as the η distribution,simulation provides a reasonable description of the data.

1 The primary vertex is identified as the reconstructed vertex with the highest sum of track pT associatedto it.

47

48 M E A S U R I N G T H E P E R F O R M A N C E O F T H E P I X E L C L U S T E R N E U R A L N E T W O R K I N D ATA

Eve

nts

/ 0.5

GeV

310

410

510

610

= 13 TeVsATLAS Preliminary

Leading track (Data)Leading track (di-jet MC)Sub-leading track (Data)Sub-leading track (di-jet MC)

[GeV]T

p0 5 10 15 20 25 30 35 40 45 50

Dat

a / M

C

0.60.8

11.21.4

Figure 27: Comparison of leading and sub-leading track pT distributions of selected track pairsin data and MC simulation. The simulation is normalized to match the number oftrack pairs in data.

8.1.1 Extrapolation Measurement

Collimated tracks from prompt particles are most likely to create merged clusters in theinnermost layers. For this reason, only results for the IBL and the B-layer are presentedin the following. They are obtained as follows:

• A track pair, with associated clusters in the B-layer or 3rd pixel layer is usedas a starting point to measure the properties of clusters in the IBL or B-layerrespectively.

• The position of the particles at the IBL or B-layer is obtained by extrapolatingback along the reconstructed trajectory of the particle, while taking into accountpossible effects due to multiple scattering.

• If a cluster is found near the extrapolated position on the inner layer, the separa-tion between the extrapolated position and that clusters is used as an estimateof the distance of the pair of tracks at that layer.

• Depending if the track reconstruction labeled the cluster as shareable, the foundcluster is categorized as either being a

1. single-particle-cluster,

2. shareable cluster (used by both tracks),

3. shareable cluster (used by only one track).

In principle, the accuracy of the extrapolation is known to be very high. Neverthe-less, due to residual mis-alignments of the pixel sensors resulting from inaccuratelyknown positions of the pixel modules, a systematic uncertainty is introduced. Thisuncertainty is not quantified explicitly, but qualitatively monitored with the comple-mentary measurement described in Section 8.1.2.

8.1 M E T H O D S 49

Single-particle-clusters created and used by only one track are expected to constitutethe majority, and most of the clusters are identified as such. If the ambiguity solver,utilizing the NN pixel clustering, tags a cluster as shareable and this cluster is usedby both tracks, it will end up in the second category. In the last category, a cluster isidentified as being shareable, but only one track is associated to it. For these clusters,the second track of the pair must have a cluster on the same pixel module associatedto it. All three categories are illustrated in Figure 28.

The fraction of clusters in the first category is expected to decrease with decreasingseparation between the tracks. At a small enough separation, on the order of the sizeof a single pixel, merged clusters are unavoidable. With an optimally performant re-construction, a population inversion should be produced, and all clusters would endup in the second category for shareable clusters. Category 3 can point to imperfectionsin the algorithmic performance of the ambiguity solver and the NN pixel clustering,but can also arise from shortcomings of the overall track reconstruction. One examplewould be errors in the combinatorial track finding described in Section 4.2. Clustersin this category do not necessarily lead to a decreased track reconstruction efficiencyor an increased track resolution. They are therefore merged with Category 1 for thismethod.

Figure 29 (a) shows a schematic of this method.

8.1.2 Measurement Using Overlap Regions

The fact that the pixel modules in the individual layers overlap can be exploited toprobe the identification power of merged clusters of the track reconstruction. All pixel(IBL) staves are installed with a specific tilt of 20° (25.714°) - where the angle describesthe tangent to the plane perpendicular to the cylinder axis. An angular overlap in theφ-direction is created, which for the IBL is 1.821°. For all layers except the IBL anoverlap between modules in the z-direction also exists.

Particles passing through these overlap regions are likely to create clusters on twomodules in the same layer. Similar to the method outlined in Section 8.1.1, the outercluster can be used as a reference and the inner one can again be assigned to one of thethree categories. In contrast to the extrapolation method, no extrapolation is necessaryfor these topologies. Since the two clusters are nearly at the same radius, the distanceof the two tracks at the inner module can just be taken as the distance of the two atthe outer one. Since the relative position of the two modules is known with a highaccuracy, the impact of alignment related uncertainties on this method is insignificant.Furthermore, due to the small radial distance between the two modules, the “type” ofthe two clusters will be strongly correlated, such that if the outer cluster is a mergedcluster, the inner cluster should most likely also be merged and the track reconstructionshould identify it as such. Conversely, if the outer cluster is not identified as beingmerged, one can probe for false-positives if the inner cluster is identified as such. Anydiscrepancy in the categorization of the two clusters can indicate inefficiencies of thetrack reconstruction. An incorrect prediction by the NN pixel clustering will be by farthe dominant source of such an inefficiency. Other sources, like wrongly assigning anearby cluster to the track also exist. Figure 29 (b) sketches how this method worksand Figure 30 gives a schematic view of the scenarios described above.


Single-particle-cluster

(a)

Shareable cluster (used

by both tracks)

(b)

Single-particle-

cluster Shareable cluster (used by only one

track)

(c)

Figure 28: Illustration of the different cluster categories:: (a) single-particle-cluster, (b) share-able cluster (used by both tracks) (c) shareable cluster (used by only one track).Both clusters have to be on the same module.

8.2 R E S U LT S 51

IBL

B-layer

Extrapolate into inner layer

Then, check cluster exists

(a)

IBL

B-layer

Then, look for the other cluster in the same layer (different module)

Ask cluster as reference

(b)

Figure 29: Sketch of (a) the extrapolation method and (b) the method using the overlap re-gions.

8.2 R E S U LT S

A qualitative agreement between data and MC simulation for the distance of two tracksof a pair at the IBL and B-layer is essential to interpret the results. Figure 31 showsthat this distance is well modeled in MC simulation, with residual discrepancies of 10to 20% at very small track separations. For the selected subset of tracks, the fractionof clusters being identified as shareable is 5.9% and 5.6% on the IBL and B-layerrespectively.

Figure 32 shows the fraction of track pairs in each category as a function of thedistance between two tracks. As previously explained, clusters of Category 1 and 3are merged and called single-particle clusters in these results. As the separation be-tween the two tracks decreases, the fraction of clusters being identified as shareableincreases. Data agrees with MC simulation to within a few percent. Below a separationof the order of two single pixels (about 0.1 mm) clusters must start to merge, and apopulation inversion is observed. For the IBL (B-layer) the fraction of clusters identi-fied as being shareable and assigned to both tracks reaches approximately 85% (93%)for data and simulation.

Due to imperfect modeling of pixel clusters in simulation, certain discrepancies areexpected. For example, a simplified charge deposition model is used in simulationwhich is known to imprecisely describe the data. Furthermore, the constant radiationcreated by the LHC’s collisions and interacting with the sensors changes the observedcluster properties with time in data, something which is not accounted for in simu-lation. The observed overall discrepancy in the fraction of cluster of Category 2 forthe B-layer of about 10% is therefore expected. Due to the smaller longitudinal pixeldimensions in the IBL the overall higher fraction of clusters of Category 1 comparedto the B-layer is reasonable.


Target layer:

Single-particle-

cluster


by both tracks)

Reference layer:

Single-particle-cluster

Shareable cluster (used by

only one track)

(a)

Target layer:Single-particle-

cluster Shareable cluster (used by both tracks) Shareable cluster (used by only one

track)

Reference layer:


by both tracks)

(b)

Target layer:

Single-particle-

cluster Shareable cluster (used by both tracks) Shareable cluster (used by only one

track)

Reference layer:

Single-particle-cluster Shareable cluster (used

by only one track)

(c)

Figure 30: Illustration of the different cluster categories: (a) single-particle-cluster, (b) share-able cluster (used by both tracks) (c) shareable cluster (used by only one track), onthe reference layer. For each case, the cluster on the target layer can also be fromone of the three categories.

8.2 R E S U LT S 53

Eve

nts

/ 0.1

mm

0

20

40

60

80

100

120

310×

ATLAS Preliminary

= 13 TeVs

Data

Di-jet MC

Measurement using extrapolation

IBL

Distance of tracks at layer [mm]0 0.5 1 1.5 2 2.5 3 3.5 4

Dat

a / M

C

0.60.8

11.21.4

(a)

Eve

nts

/ 0.1

mm

0

20

40

60

80

100

310×

ATLAS Preliminary

= 13 TeVs

Data

Di-jet MC


B-layer

Distance of tracks at layer [mm]0 0.5 1 1.5 2 2.5 3 3.5 4

Dat

a / M

C

0.60.8

11.21.4

(b)

Figure 31: Distance between track pairs for (a) IBL and (b) B-layer, respectively in data andMC simulation.

Another interesting observation is that the fraction of Category 2 clusters is smallerfor track distances below ∼0.1 mm for the IBL than for the B-layer. This hints at aworse performance of the NN clustering in that layer. As pointed out in Section 3.4.1,this is due to the fact that the IBL only provides a 4-bit readout compared to the 8-bitreadout used for the other pixel layers.

Distributions of the fraction of clusters in each category, from the method using theoverlap region, are presented in Figure 33. Clusters from categories 1 and 3 are againmerged in this figure. Compared to Figure 32, the overall features observed in alldistributions are the same. For the B-layer, the fraction of clusters identified as beingshareable and assigned to both tracks is approximately 95% and 90% for data andsimulation, respectively. Clusters on the IBL again show a significantly lower efficiencyas on the B-layer, about 80% in both data and simulation, while the overall agreementbetween data and simulation is within 10%. Figure 34 shows the fraction of clusterson the inner module with a different category with respect to the cluster in the outermodule. For this figure the three categories are considered separately. For increasinglysmall separations the fraction of clusters with mismatching categories increases. Dueto the smaller pixel dimensions of the IBL, this increase occurs at smaller separationsthan for the B-layer. At separations where clusters must merge (around 0.1 mm), thereis a drop in the distribution for both the IBL and B-layer. Mismatches in the categoriescan arise for a variety of reasons. As described in Section 4.3, the NN pixel clusteringcan fail to identify a cluster as being merged, or wrongly identify it as being so. Therewill always be fluctuations in the shape of the clusters, even with identical incidentangles. Therefore, there will never be a perfect match between the categories of thetwo clusters on the same layer, even if simulation would perfectly describe data.


Fra

ctio

n of

clu

ster

cla

ss /

0.1

mm

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Tracks with single-particle-cluster:

Data

Di-jet MC

Tracks with shareable cluster:

Data

Di-jet MC

ATLAS Preliminary = 13 TeVs


IBL

Distance of tracks at layer [mm]0 0.2 0.4 0.6 0.8 1 1.2

Dat

a / M

C

0.60.8

11.21.4

(a)

Fra

ctio

n of

clu

ster

cla

ss /

0.1

mm

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1


Data

Di-jet MC


Data

Di-jet MC



B-layer


Dat

a / M

C

0.60.8

11.21.4

(b)

Figure 32: Fraction of tracks with single-particle and shareable clusters as a function of thedistance between tracks using the extrapolation method for (a) IBL and (b) B-layer.

Fra

ctio

n of

clu

ster

cla

ss /

0.1

mm

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1


Data

Di-jet MC


Data

Di-jet MC


Measurement using overlap region

IBL


Dat

a / M

C

0.60.8

11.21.4

(a)

Fra

ctio

n of

clu

ster

cla

ss /

0.1

mm

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1


Data

Di-jet MC


Data

Di-jet MC



B-layer


Dat

a / M

C

0.60.8

11.21.4

(b)

Figure 33: Fraction of tracks with single-particle and shareable clusters as a function of thedistance between tracks using the overlap method for (a) IBL and (b) B-layer.

8.3 C O N C L U S I O N 55

Fra

ctio

n of

diff

eren

t clu

ster

cla

ss /

0.1

mm

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

Cluster on IBL:

Data

Di-jet MC

Cluster on B-layerData

Di-jet MC




Dat

a / M

C

0.60.8

11.21.4

(a)

Figure 34: The fraction of tracks with mismatching categories between reference and targetcluster.


Studying the properties of pixel clusters for collimated track pairs allows the probingof the identification efficiency of the track reconstruction of merged pixel clusters indata. For separations between the pair of tracks below the dimension of a single pixel,the efficiency for identifying the cluster as shareable and assigning it correctly to bothtracks is above 80% for the IBL and above 90% for the B-layer. The performance ofthe track reconstruction is consistent between data and simulation, with residual dis-crepancies of 5% and 10% for the extrapolation and overlap methods respectively.

9M E A S U R I N G T R A C K L O S S I N D E N S E E N V I R O N M E N T S W I T H T H ED E /D X M E T H O D

Detailed understanding of uncertainties related to the track reconstruction perfor-mance in dense environments is crucial for several performance studies, includingcalibration of the jet energy scale (JES) using charged particle quantities [35] andcalibration of the jet mass in large-radius jets (see Chapter 10). These uncertaintiescan also directly affect analyses, which is discussed in the context of this thesis inSection 11.6.

Previous studies showed that the leading source of uncertainty was the track recon-struction efficiency in dense environments [48]. As described in Section 5, updatesto the track reconstruction setup significantly improved its efficiency. Still, residualinefficiencies exist (see Chapter 6).

This chapter presents a novel data-driven method which uses the measure of ioniza-tion energy loss (dE/dx) provided by the pixel detector to quantify this track recon-struction inefficiency. Using the fact that a single charged particle produced in colli-sions at the LHC is expected to be a minimum ionizing particle (MIP), one can use thedE/dx in pixel clusters to distinguish clusters created by one or two MIPs. This can beused to identify the residual inefficiencies by asking how often are there two recon-structed tracks associated to a cluster from two particles. By comparing the measuredinefficiency in data and MC, the respective uncertainty on the track reconstructionefficiency can be estimated.

As the previous chapters have shown, tracks have the highest collimation closest tothe interaction point. Therefore the probability of finding a merged cluster is greaterfor the inner layers of the pixel detector. Clusters on the IBL cannot be used for thisstudy since it does not measure dE/dx with sufficient resolution. Again, this is a resultof its 4-bit ToT resolution, compared to eight bits used for the remaining pixel layers.As a result, this study uses information from clusters on the B-layer.

9.1 M E T H O D

The reconstruction and calibration of the R = 0.4 jets used in this study are describedin Section 6.1. The same data set and simulated samples as described in Chapter 7 areused.

Reconstructed tracks must fulfill the following requirements:

• exactly one pixel cluster per layer

• pT > 10 GeV

• |η|< 1.2

• |dBL0 |< 1.5 mm

• |zBL0 sinθ |< 1.5 mm

57

58 M E A S U R I N G T R A C K L O S S I N D E N S E E N V I R O N M E N T S W I T H T H E D E/D X M E T H O D

• number of SCT clusters ≥ 6

where dBL0 is the transverse impact parameter calculated with respect to the mea-

sured beam line position, zBL0 is the difference between the longitudinal position of

the track along the beam line at the point where dBL0 is measured and the longitudinal

position of the primary vertex, and θ is the polar angle of the track. These stringentrequirements select topologies with enhanced contributions of high quality collimatedtracks, with negligible contributions from fake tracks.

Using data alone, this method shall quantify residual inefficiencies of the track re-construction in the core of jets. To achieve this, three distinct distributions of clusterdE/dx are created through the selections outlined in Figure 35.

<0.05

JetpT

cut

Multiply Used

Inside Jet Core

Multiple-track template

Sample

Event Selection Jet Selection

Track Selection

0 < l11l < 1.2 Trackp

T > 10 GeV

1 pixel hit per layer B-layer dE/dx

Not Multiply Used

Inside Jet Core

Target

distribution

Outside Jet Core

Single-track template

AR{jet,track) >0.1

AR{jet,track)

Figure 35: Selection chart for the three distinct dE/dx distributions used in the analysis.

One of them represents the dE/dx distribution of clusters from one MIP. This isachieved by selecting tracks which are outside the core of the jet (∆R(jet,track)> 0.1)and which use a cluster in the B-layer which is not used on any other track. For suchtracks it is very likely that their clusters are created by one single particle and arelabeled Not-Multiply-Used in the following. The second dE/dx distribution is built fromtracks which are inside the core of the jet (∆R(jet,track)< 0.05) and which contain acluster in the B-layer which is also used by another track. These clusters have a highprobability to be created by more than one particle, i.e. to be merged clusters, and arefurther defined as Multiply-Used.

Figure 36 shows the two dE/dx distributions. The Not-Multiply-Used distribution,plotted as blue circles, exhibits a single peak at the dE/dx value expected for a MIP. TheMultiply-Used distribution, plotted as green squares, instead peaks at a dE/dx valueexpected for two MIPs. At dE/dx> 3.2 MeVg−1cm2 one can also identify further peaks,which corresponds to B-layer clusters created by more than two charged particles. SuchMultiply-Used clusters are still possible as was explained in Section 4.3.

The final distribution, the so called target distribution, is created from tracks insidethe core of the jet (∆R(jet,track)< 0.05), which have a B-layer cluster which is onlyassociated to one track. This subset of pixel clusters has two major populations, oneconsisting of clusters created by exactly one particle and correctly assigned to exactly

9.1 M E T H O D 59

one track, and one of merged clusters created by multiple particles, which are onlyassigned to one track due to track reconstruction inefficiencies. Using the first twodE/dx distributions of Not-Multiply-Used and Multiply-Used clusters as templates, onecan fit the target distribution to resolve the fractions of the two populations. With theresulting fitted fraction of merged clusters, Fmerged, with only one track associated tothem, one can set an upper limit on the fraction of unreconstructed particles whichhave created merged clusters:

F lost =NLost

NTrue2

=NLost

NReco2 + 2 ·NLost

, NLost = Fmerged ·NRecoData , (5)

where NLost is the number of lost tracks, NTrue2 is the true number of tracks that

should be associated to a merged B-layer cluster from two particles, NReco2 is the num-

ber of reconstructed tracks inside the multiple-track template and NRecoData the number of

reconstructed tracks in the target distribution. In the denominator, the number of losttracks needs to be counted twice, since for every measured lost track two tracks aremissing from the multiple-track template. The second equality is only strictly true ifthe assumption that lost tracks have the same properties as reconstructed tracks holds,which is known to be true from simulation. In other words, the number of tracks asso-ciated to a cluster created by two charged particles which would fail the track selectionis negligible.

Since the peak in the dE/dx distribution from two MIPs for Multiply-Used clustersis by far the most prominent and clean, the fit is performed only between 1.1-3.07(1.26-3.2) MeVg−1cm2 for data (simulation) to suppress contributions from higherdE/dx values. A imperfect description of the leading edge of the target distributionby the single-track-template would affect the fitted result. Since the area of interestlies at much higher dE/dx values, the lower edge of the fit range was chosen to avoidas much as possible the leading edge of the one particle dE/dx peak, while retaininghigh statistics for the remainder of the distribution. Simulation-based studies showedthat no significant bias of F lost is induced due to the choice of fit range, i.e. the truevalue of F lost for tracks from clusters from the restricted and the inclusive range arecomparable.

F lost measures possible inefficiencies for tracks from particles using merged clustersfrom exactly two particles. The fit range is adjusted in simulation in order to have thesame fraction of clusters as in the target distribution, which are inside the fitted rangewith respect to the total number of clusters. This is necessary since the calibratedresponse of the pixel sensors to a specific energy deposit has certain imperfectionswhich results in a shift of the overall dE/dx distribution.

The used fitter fits an input data histogram (d) with two templates (T1 and T2) usinga standard binned log likelihood fit and two fit parameters (n and Fmerged) with thefunction:

f (x , Fmerged, n) = n× [(1− Fmerged)× T1(x)+ Fmerged× T2(x)] (6)

where T1/2(x) is the number of entries in the template 1 and 2 in bin x , n is a floatingnormalization factor and Fmerged is the fit parameter for the fraction of template T2

(Multiply-Used template) entries in the data in bin x . For each bin x in the fit range

]2 cm1dE/dx [MeV g

0 0.5 1 1.5 2 2.5 3 3.5 4

Entr

ies

0

5000

10000

15000

20000

25000

30000

35000

40000

< 400 GeVjet

T200 GeV < p

, s = 13 TeV, 3.2 fb-1

SingleTrack Template

MultipleTrack Template

Figure 36: dE/dx distributions of the single- and multiple-track templates for data with a jetpT in the range 200 GeV < p jet

T < 400 GeV.

the compatibility of the observed data with the fit model prediction is computed usinga Poisson likelihood. The best set of parameters is found by minimizing the negativelogarithm of the product of the Poisson likelihood over all bins in the fit range:∑

x

d(x)× log f (x , Fmerged, n)− f (x , Fmerged, n) (7)

Studies on the single ρ sample, also used in Chapter 5, show that the dE/dx distri-bution obtained by the selections applied to the target distribution is consistent withselecting potentially lost tracks. This reinforces the assumption that the fitted fractionFmerged corresponds to the fraction of lost tracks. For a single ρ decaying into two pi-ons, two distinct event topologies can be observed. Firstly, the number of reconstructedtracks can match the number of reconstructed pseudo-tracks, and secondly, only onetrack can be reconstructed while two pseudo-tracks exist.

Figure 37 shows the dE/dx distributions for the two cases, where the distributionwith one reconstructed track but two pseudo-tracks represents the dE/dx distributionwhen one track of a pair of collimated tracks is lost. For this distribution, the domi-nating second peak is due to the merged pixel clusters. The peak at a dE/dx of oneMIP corresponds to tracks lost due to other effects than merging of the B-layer cluster.To check if the selections applied to obtain the Multiply-Used template really corre-spond to this distribution of lost tracks, the same selections are applied to this sample.The result can be seen in Figure 38, where the Multiply-Used template matches thedistribution observed for lost tracks.

The jet pT is directly related to the density of the environment and the frequencyof merged clusters, and therefore expected to change the fraction of charged particleswhich might not be reconstructed into a track due to potential inefficiencies. For thisreason the study is performed in jet pT bins of 200 GeV ranging from 200 GeV to1600 GeV. For simulation, jet pT bin dependent templates are used. In the highest jetpT bins in data, statistics is low. To not impair the fit quality due to the low statistics,the single-track and multiple-track templates obtained from the 200-400 GeV bin are

9.2 S Y S T E M AT I C U N C E RTA I N T I E S 61

Figure 37: dE/dx distributions for two possible event topologies in the single ρ MC sample.The first topology with one reconstructed track but two pseudo-tracks representsthe dE/dx distribution when one track of a pair of collimated tracks is lost. Thesecond topology corresponds to the case where both tracks are reconstructed.

used to fit all jet pT bins in data. Within the statistical uncertainty, these templateshave the same shape as the ones at higher jet pT within the fit range.

9.2 S Y S T E M AT I C U N C E RTA I N T I E S

Two different groups of systematic uncertainties need to be considered. The first areuncertainties originating from the methodology described in Section 9.1. The secondgroup arises from differences in the physics modeling between several MC event gen-erators, and only affect the final data and MC comparison.

9.2.1 Uncertainties Related to the Method

Clusters created by more than two particles have a non-negligible contribution to thetarget dE/dx distribution. The choice of fit range, especially the upper bound, is ex-pected to affect F lost, the measured fraction of merged clusters associated to only onetrack. This effect is studied by increasing the fit range at higher dE/dx above the base-line selection. With each increment of 0.2 MeVg−1cm2, the relative increase in F lost

is about 5%. The maximum change in F lostfor each bin of jet pT is added as an sym-metrized uncertainty in that bin.

Due to the fitting of data in all jet pT bins with templates from the lowest jet pT bin,an additional systematic uncertainty is introduced. From studies on MC simulation,this procedure is known to introduce a small systematic bias in the fitted Fmerged. This iscaused by the varying contamination from clusters created by more than two particlesin the multiple-track template with jet pT, which varies between 3% and 8%. A jet pT

dependent correction factor is therefore applied to the result in data and an uncertaintyon the order of the difference observed in simulation is assigned.


Figure 38: Distributions of dE/dx for events in the single ρMC sample where only one track ofthe decay products of the ρ is reconstructed, but both are expected to be, comparedto dE/dx distribution obtained in the Multiply-Used template.

Another possible source of uncertainty are possible contributions to F lost not origi-nating from the density of the environment. Such contributions could come from pile-up tracks creating merged clusters with tracks in the jets, as well as isolated track beinglost. Conservative estimates based on truth-level studies on MC simulation showed thatsuch contributions are on the order of 2% to 6% of the total Fmerged in the studied jetpT range, so much smaller as the other systematic uncertainties, and therefore are notconsidered explicitly in the final results. However, residual non-closure is consideredinclusively.

A truth-based closure test was performed on simulated samples to validate themethod for any missed systematic bias or contribution. It was found that in the lowestjet pT bin, a significant non-closure is observed of the order of 18% - this correspondsto an absolute difference in the measured F lost of about 0.013. This non-closure quicklydecreases in the higher jet pT bins. The non-closure arises mainly from two sources:the residual dE/dx peak at values expected from one MIP in the multiple-track tem-plate, and large dE/dx values from more than two clusters. The former effect is morerelevant at low jet pT while the second effect becomes stronger at high jet pT. Thenon-closure is added as an additional uncertainty in all jet pT bins for both simulationand data. Table 1 lists the relative uncertainties on F lost.

Uncertainties on the JES calibration or resolution have negligible impact on theanalysis. No effects due to the choice of binning in the individual bins of jet pT are ex-pected since the bin size is much smaller than the dE/dx resolution. This was checkedexplicitly by varying the bin size around the default binning used in the analysis. Allvariations produce identical results to the nominal binning. All relative systematic un-certainties on F lost are summarized in Table 1.

9.3 R E S U LT S 63

Table 1: Summary of the relative systematic uncertainties on the fraction of lost tracks (F lost)in bins of jet pT for data.

Jet pT bin Fit Range [%] Low pT Templates [%] Non-closure [%]

200–400 GeV 13 0 18

400–600 GeV 12 7 11

600–800 GeV 10 13 6

800–1000 GeV 12 18 1

1000–1200 GeV 12 21 0

1200–1400 GeV 11 16 0

1400–1600 GeV 15 16 0

9.2.2 Generator Uncertainties

In order to study the overall agreement of MC simulation with data, one has to takedifferences between different simulated models into account. Generator dependent dif-ferences provide the dominant systematic uncertainty. Comparing the measured valuesof Fmerged for samples generated with either PYTHIA 8, SHERPA [49] or Herwig++ [50]allows to estimate this uncertainty as is shown in Figure 39. The observed differencesare most likely due to different modeling of the charged particle multiplicities anddistributions inside jets. For each bin of jet pT, the assigned uncertainty is the largestsymmetrized difference between the fitted value of Fmerged between either of the threegenerators. This results in relative uncertainties of Fmerged between 4 and 34%. It isimportant to note that this uncertainty does not need to be considered if one onlycompares a specific simulated sample to data.

9.3 R E S U LT S

Target distributions for data, fitted with the Not-Multiply-Used and Multiply-Used tem-plates, are shown in Figure 40 for two bins of jet pT. The combinations of both tem-plates are able to well describe the target distribution, since the ratio of the fittedfunction to the target distribution is consistent with unity.

Figure 41 shows that F lost is slightly increasing with jet pT and that simulation isable to describe the data. In the core of jets, the fraction of lost tracks ranges from0.061± 0.006(stat.)± 0.014(syst.) to 0.093± 0.017(stat.)± 0.021(syst.) between ajet pT of 200 to 400 GeV and 1400 to 1600 GeV, respectively. This increase is causedby increasingly boosted particles, which cause higher collimation of the track pair, andis not due to confusion in correctly assigning clusters to tracks. At a certain point, thetwo particles are so collimated that the reconstructed tracks start to overlap completelyup to the radius of the SCT detector. At that point a similar effect as shown for tracksfrom the ρ decay in Figures 12 and 15 occurs. The cluster assignment efficiency for re-constructed tracks remains constant with increasing jet pT, indicating no degradation


[GeV]T

Jet p200 400 600 800 1000 1200 1400 1600

mer

ged

F

0

0.005

0.01

0.015

0.02

0.025

0.03

Simulation, s = 13 TeV

Pythia8Herwig++Sherpa

Figure 39: Comparison of the fitted fraction of merged clusters, Fmerged, between Pythia8, Her-wig++, and Sherpa as a function of jet pT. Errors bars are purely statistical whilethe red error band for Pythia8 reflects the derived generator uncertainty.

of performance due to the environmental effects besides the second track. The prob-ability of loosing one of the tracks rises only because of their increasing collimation.This effect has been confirmed on simulation for tracks selected by this analysis.

Fitting a constant value to the ratio of F lost shows that data and MC simulation areconsistent across the whole studied jet pT range.

9.3 R E S U LT S 65

Entr

ies

20

40

60

80

100

120

310×

< 400 GeVjet

T200 GeV < p

0.001 (stat)± = 0.008 merged

F

Data

Fit

SingleTrack Contribution MultipleTrack Contribution

]2 cm1dE/dx [MeV g

0 0.5 1 1.5 2 2.5 3 3.5 4

Fit/D

ata

0.80.9

11.11.2

, s = 13 TeV, 3.2 fb-1

(a)

Entr

ies

200

400

600

800

1000

1200

1400

, s = 13 TeV, 3.2 fb-1

< 1200 GeVjet

T1000 GeV < p

0.005 (stat)± = 0.028 merged

F

Data

Fit

SingleTrack Contribution MultipleTrack Contribution

]2 cm1dE/dx [MeV g

0 0.5 1 1.5 2 2.5 3 3.5 4

Fit/D

ata

0.80.9

11.11.2

(b)

Figure 40: Comparison of target dE/dx distributions with the fitted results for a jet pT between(a) 200-400 GeV and (b) 1000-1200 GeV. The two template distributions scaledby their fitted fractions are also shown. The ratio between the fit and the targetdistribution is shown on the bottom within the fit range (1.1–3.07 MeVg−1cm2).

[GeV]T

Jet p

200 400 600 800 1000 1200 1400 1600

lost

F

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

s=13 TeV, 3.2 fb-1

Data

Simulation (Pythia 8)

Figure 41: The measured fraction of lost tracks (F lost) in the core of jets (∆R(jet,track)< 0.05)as a function of jet pT for data and simulation. The gray (red) error bands indi-cate the total uncertainty, while the black (red) error bars show only the statisticaluncertainty for data (simulation).

Part IV

J E T M A S S R E C O N S T R U C T I O N

10J E T M A S S R E C O N S T R U C T I O N

Massive particles produced in the decay of TeV-scale resonances, such as W and Zbosons or top quarks, are highly boosted. As a result, their decay products becomeincreasingly collimated as the mass of the resonance increase. In case they decay intoquarks, they are reconstructed as a single large-radius jet. One of the strongest dis-criminants between these jets and the multijet background is their mass (jet mass).Improving the resolution (JMR) of the reconstructed jet mass and reducing relateduncertainties is therefore essential.

This part of the thesis introduces the different approaches to jet mass reconstructionused in ATLAS. Most of the content in this chapter has been previously presented inReference [51] with significant contributions from the author of this thesis. Where notexplicitly stated, all result in this thesis are based on the optimized track reconstructionsetup discussed in detail in Part II. MC samples are used in this chapter to obtain theexpected performance of multiple jet mass definitions, to perform the mass calibration,as well as for the derivation of related uncertainties. The same MC and data samplesas described in Chapter 7 are used. In addition, simulated samples of W ′ boson (Z ′

boson) event provide a rich source for boosted W /Z bosons and top quarks. They aresimulated with the same event generator (PYTHIA), the same PDF (NNPDF23LO) andtuned parameters (A14) as the dijet MC samples.

Two basic jet mass definitions, each with strength in specific kinematic regimes, anda method to combine the two are presented. Built solely from calorimeter information,the calorimeter-based jet mass serves as the baseline mass definition. This mass defi-nition can be expanded, using the higher granularity information available from theID. This leads to the definition of the so called track-assisted jet mass. By combiningthese two variables, one can achieve the performant combined jet mass. Each of thesejet mass definitions are described in more detail in Sections 10.1 and 10.5. After intro-ducing the various mass definitions, their resolutions are compared in simulation inSection 10.3 and corresponding systematic uncertainties are discussed in Section 10.4.

10.1 J E T M A S S D E F I N I T I O N S

The jet mass is mainly used as a variable to identify jets originating from boostedparticles. For these topologies large-radius jets are used and the following results arepresented for such jets. However, the general mass definitions provided below, as wellas their calibration procedures, are also valid for small-radius jets.

The classical jet mass definition within ATLAS is the calorimeter-based jet mass(mcalo). For a calorimeter jet J composed of calorimeter-cell clusters i with energyEi , momentum ~pi (|~pi|= Ei) it is defined as:

69

70 J E T M A S S R E C O N S T R U C T I O N

mcalo =

√

√

√

√

�

∑

i∈J

Ei

�2

−

�

∑

i∈J

~pi

�2

. (8)

Since this jet mass depends on both angular (used to build ~pi) and energy informa-tion, the resolution of both individual measurements will determine the resolution ofthe mass. As described in Section 3.5, the calorimeters offer excellent energy resolu-tion. However, the spatial granularity, which determines the angular resolution, canbecome a limiting factor for very collimated streams of particles. For a sufficiently highLorentz boost of the parent particle, the spatial separation of its decay products is sosmall that their spatial separation reaches the calorimeter granularity. In the limit ofa jet reconstructed from a single calorimeter-cell cluster, the jet mass following theclassical definition above becomes zero.

As a way out, high resolution angular information from the tracking detectors canbe used to complement the calorimeter measurements. Similar approaches have pre-viously shown to provide superior performance, for example in particle-flow [52].

There have been previous attempts to improve the reconstructed jet mass, the first ofwhich extended the electromagnetic calorimeter measurement by using the hadronicinformation [53, 54]. In the context of tagging jets originating from a top quark, infor-mation from the tracker was used for the first time to complement the calorimeter mea-surement of the mass [55]. Phenomenological studies for future colliders discussed atrack-assisted mass in connection with highly boosted (pT ¦ O (10) TeV) jets frombosons and top quarks [56, 57].

Information from the tracker and the calorimeter can be combined into the track-assisted jet mass (mTA):

mTA =pcalo

T

ptrackT

×mtrack, (9)

where pcaloT is the transverse momentum of a calorimeter jet, ptrack

T is the transversemomentum of the sum of all tracks’ four-vectors associated to the jet, and mtrack isthe invariant mass of this sum. For a single reconstructed track the mass is set tomπ, a good approximation since ptrack

T � mπ . The ratio of the transverse momentapcalo

T /ptrackT corrects for the component from neutral particles, which the ID cannot

reconstruct, and which is missing in ptrackT . This minimizes the effect of charged-to-

neutral fluctuations, and improves the resolution with respect to a track-only jet massdefinition (mtrack). Figure 42 compares the uncalibrated calorimeter-based jet mass(dashed red line) to the uncalibrated track-assisted jet mass (dashed black line), andshows that the latter is superior in resolution with respect to the uncalibrated track-only jet mass (dashed blue line) for jets with 1.6 TeV< pT < 1.8 TeV.

Besides the charged-to-neutral fluctuations, the resolution and purity of the subsetof tracks selected as input to the track-assisted mass limits its resolution. By replacingthe reconstructed tracks directly with the charged particles associated to the truthjet matched to the calo jet, one can estimate the ultimate possible performance ofthis method using MC simulation. As shown in Figure 43, the resolution would be

10.2 J E T M A S S C A L I B R AT I O N 71

Jet mass [GeV]

0 50 100 150 200

Fra

ctio

n /

4 G

eV

0

0.05

0.1

0.15

0.2

0.25 Simulation PreliminaryATLAS

= 13 TeV, W/Z-jetss

| < 0.4η < 1.8 TeV,|T

1.6 TeV < p

calomtrackmTAm

UncalibratedCalibrated

Figure 42: Reconstructed jet mass distributions for calorimeter-based jet mass, mcalo, track-assisted jet mass mTA and the track-only jet mass mtrack for W /Z-jets. Each mass isshown before and after calibration.

improved by a factor of two, if the efficiency to reconstruct all tracks would be 100%,their resolution zero, as well as the assignment of the tracks to the reconstructed jetwould be perfect.

truth / mTAm

0 0.5 1 1.5 2

Fra

ctio

n / 0

.1

0

0.2

0.4


= 13 TeV, W/Z-jetss

< 90 GeVtruth80 GeV < m

| < 0.8η < 2.4 TeV, |truth

T2.2 TeV 30 ps (excluding muons)as input. Jets are restricted to 2.2 TeV< pT < 2.4 TeV and 80 GeV< mtruth < 90 GeV.

10.2 J E T M A S S C A L I B R AT I O N

Analogously to the calibration applied to correct the JES [35, 37, 38], also the jetmass scale (JMS) needs to be calibrated. The goal of this calibration is to correct thereconstructed jet mass on average to the mass of the jet on truth-particle-level.

The mass response for a given jet mass is defined as the ratio between the recon-structed jet mass and the mass of the matched truth-level jet:

Rm = mreco/mtruth; mreco ∈ cJES ·mcalo, cJES ·mTA, (10)

where cJES is the Ereco and ηdet dependent JES calibration factor - hence the masscalibration is performed with the JES calibration already applied.

The jet mass calibration factors are determined by the mean of a Gaussian fit to theaverage response ⟨Rm⟩ in bins of ptruth

T , |ηdet|, mtruth. This calibration is performed ina sample of QCD dijet MC by matching isolated large-radius jets to isolated truth-leveljets. Isolation requires no other jet within∆R = 1.5 (2.5) for the reconstructed (truth)jet. The truth-level jet is matched if its pT > 100 GeV and it is within ∆R < 0.6 of thereconstructed jet. To transform the dependence of this calibration from truth basedvalues to reconstructed jet quantities, a numerical inversion is performed [51]. Theaverage jet mass response is said to be calibrated if ⟨R⟩= 1.

Although the JES calibration partially corrects for η dependent detector effects, afull jet mass calibration is still required. Figure 44 shows the average response versusptruth

T after energy, but before mass calibration. A much more uniform response beforemass calibration for mTA is observed compared to mcalo, since pcalo

T in Equation 9 isalready calibrated.

[GeV]T

Truth jet p0 1000 2000 3000

⟩ tr

uth

/ m

calo

m⟨

1

1.5

2 Simulation PreliminaryATLAS

| < 0.4det

η = 13 TeV, QCD dijets, |s

Energy-calibration only

< 60 GeVtruth40 GeV < m < 100 GeVtruth80 GeV < m < 200 GeVtruth160 GeV < m

(a)

[GeV]T

Truth jet p0 1000 2000 3000

⟩ tr

uth

/ m

TA

m⟨

1

1.5


| < 0.4det


Energy-calibration only


(b)

Figure 44: Average jet mass response of (a) mcalo and (b) mTA as a function of ptruthT for central

jets in dijet MC shown in three bins of mtruth after energy but before mass calibra-tion.

As shown in Figure 45, after calibration the mcalo and the mTA have a uniform jetmass response within 3% across all bins of ptruth

T , mtruth and |ηdet|. For low mtruth andhigh ptruth

T the calibration of the calorimeter-based mass fails to achieve perfect closure.Due to the high gradient of the response of mcalo in this region, two jets on the high andlow edges of the same mtruth bin can receive very different corrections. In combinationwith the poor jet mass resolution (JMR) in this region, two populations of jets willexist: those with a higher mass where a calibration factor close to one is applied, andthose with lower mass where a large correction is applied (calibration factor � 1).This is illustrated by Figure 46 (a). For the remainder of the calibrated phase space,

10.3 J E T M A S S P E R F O R M A N C E I N S I M U L AT I O N 73

[GeV]T

Truth jet p0 1000 2000 3000

⟩ tr

uth

/ m

calo

m⟨

1

1.5


| < 0.4det


Energy and mass calibration


(a)

[GeV]T

Truth jet p0 1000 2000 3000

⟩ tr

uth

/ m

TA

m⟨

1

1.5


| < 0.4det


Energy and mass calibration


(b)

Figure 45: Average jet mass response of (a) mcalo and (b) mTA as a function of ptruthT for central

jets in dijet MC shown in three bins of mtruth after energy and mass calibration.

closure is observed after correcting the jet mass for both mass definitions, as is shownfor an example bin in Figure 46 (b).

10.3 J E T M A S S P E R F O R M A N C E I N S I M U L AT I O N

Figure 47 shows the resolution of the jet mass response as a function of truth jet pT forjets originating from boosted W and Z bosons as well as top quarks. There are severalpossibilities for quantifying the resolution of the response distribution. Using half ofthe 68% interquantile range (IQnR)1 divided by the median has proved to be a robustmeasure for this quantity, which is insensitive to outliers. In the limit of a Gaussiandistribution this quantity equals the standard deviation. Both the calorimeter-basedand track-assisted JMR degrade at high pT, as shown in Figure 47; for the calorimeter-based mass this is due to finite granularity discussed in Section 10.1 and for the track-assisted mass it is mainly due to a worsening track resolution.

For jets initiated by W or Z bosons the mTA provides superior resolution comparedto the mcalo above 1 TeV. At lower momentum the charged-to-neutral fluctuations sig-nificantly limit the resolution of the mTA so the mcalo provides the better performance.For different reasons the mTA is unable to match the performance of the mcalo for topjets in the studied phase space. Since the mass of the top is higher than the W and Zboson, a greater pT is needed to achieve the same collimation of its decay products.At sufficiently high pT a cross-over point where the mTA is superior to the mcalo is stillexpected. In addition, top jets on average exhibit a higher number of subjets than jetsfrom W or Z bosons, which again reduces the density of the environment and thereforethe granularity of the detector needed to resolve it.

1 This is defined as q84% − q16%, whereby q16% and q84% are the 16thand 84th percentiles of a givendistribution.

truth / mrecom

0 1 2 3

Fra

ctio

n / 0

.05

0

0.02

0.04

0.06


= 13 TeV, QCD dijetss


< 2.5 TeVtruth

T|<0.4, 2.25 TeV < p

detη|

calom

TAm

Uncalibrated

Calibrated

(a)

truth / mrecom

0 1 2 3

Fra

ctio

n / 0

.05

0

0.1

0.2

Simulation PreliminaryATLAS = 13 TeV, QCD dijetss


< 2.5 TeVtruth

T|<0.4, 2.25 TeV < p

detη|

calom

TAm

Uncalibrated

Calibrated

(b)

Figure 46: Jet mass response distribution for mcalo and mTA for central jets in dijet MC with2.25 TeV < ptruth

T < 2.5 TeV before and after the mass calibration is applied. Twobins of mtruth are shown: (a) 40 GeV< mtruth < 60 GeV and (b) 160 GeV< mtruth <

200 GeV.

[GeV]T

Truth jet p500 1000 1500 2000 2500

)m

) / m

edia

n(R

m 6

8% IQ

nR(R

× 21

0

0.1

0.2

0.3


| < 0.8η = 13 TeV, W/Z-jets, |s

truth/mreco = mmRcalom

TAm

(a)

[GeV]T

Truth jet p500 1000 1500 2000 2500

)m

) / m

edia

n(R

m 6

8% IQ

nR(R

× 21

0

0.1

0.2

0.3


| < 0.8η = 13 TeV, Top-jets, |s

truth/mreco = mmRcalom

TAm

(b)

Figure 47: Resolution of mcalo and mTA for central (a) W/Z-jets and (b) top-jets in MC as afunction of ptruth

T . The resolution is defined as half of the 68% interquantile range(IQnR) divided by the median of the distribution.

10.4 J E T M A S S S Y S T E M AT I C U N C E RTA I N T I E S 75

10.4 J E T M A S S S Y S T E M AT I C U N C E RTA I N T I E S

The JMS cannot be obtained from data by methods such as jet balancing, since thepartonic center of mass energy is unknown at the LHC. One alternative method forprobing the JMS in data involves the use of rm

track, defined as the ratio between mcalo

and mtrack, for QCD dijet events [38]. This quantity is proportional to the JMS underthe assumption that it can be decomposed into

rmtrack =R × (m

truth/mcharged truth)× (mcharged truth/mtrack). (11)

A residual dependence of the calorimeter response on charged-to-neutral fluctua-tions makes this factorization only approximately true. As a result,

1−⟨rmtrack⟩Data/⟨rm

track⟩MC (12)

provides an upper limit on the scale uncertainty. Uncertainties on the fragmentationmodeling, as well as the track reconstruction performance can affect the double ratioand as consequence limit the precision of this method to ∼ 5%. The method cannot beused to determine the JMR because the resolution of the track term is dominated bycharged-to-neutral fluctuations. In the case of mTA, both the uncertainties on the trackreconstruction and the calorimeter-jet pT are known, so the systematic uncertaintieson both its JMS and JMR can be evaluated directly by propagating the uncertaintyon the individual components. This is one of the key advantages of the track-assistedmass compared to the calorimeter-based mass and the uncertainties on the individualcomponents are derived as explained below.

The material in the ID is not precisely known. Since hadronic interactions with itsmaterial are the main source of track reconstruction inefficiencies for isolated parti-cles, the introduced uncertainty is evaluated by varying the amount of material withinits measured uncertainty [58]. As extensively discussed in the previous chapters, aresidual track reconstruction inefficiency in the core of highly energetic jets exists andthe uncertainties on it are derived as described in Chapter 9. Fake track provide thesecond biggest source of uncertainty on the reconstruction. Using the assumption thatthe number of reconstructed tracks is expected to increase linear with pile-up, one canestimate the number of fake tracks by measuring any deviation from this non-linearity.By comparing this non-linearity in data and MC, one obtains a 30% uncertainty on thefake rate [59]. A possible uncertainty due to a bias in the reconstructed momentum ofthe tracks can be assessed using Z → µµ events, where the momentum is reconstructediteratively to obtain the mass of the Z boson [60] (as defined by the particle datagroup [2]). As already mentioned in Section 9.2.2, the fragmentation modeling differsbetween different event generators, which leads to differences in the track reconstruc-tion performance. Such differences can lead to changes in mTA and their magnitude isestimated by comparing PYTHIA 8 and HERWIG++. The calorimeter-jet pT uncertaintyis estimated in analogue with the calorimeter based-mass: rpT

track = pcaloT /ptrack

T .The individual components of the JMS uncertainties for both mTA and mcalo are

shown in Figure 48. Full correlation is assumed between the track reconstruction un-certainties on mTA and mcalo, since for both quantities a variant of the rtrack method isused to estimate parts of the uncertainty. mTA is built through mtrack/ptrack

T so the over-all uncertainty on the JMS due to track reconstruction uncertainties is much smaller

as they cancel out in the ratio. In the region of pT = 300− 1000 GeV, the uncertaintyis about 4% for mcalo and about 2% for mTA. At higher pT, the uncertainty is limitedfor both mass definitions by the statistics available in the data set, which limits theprecision of the rtrack measurement.

[GeV]T

Jet p

210×3 310 310×2

]ca

loF

ract

iona

l JM

S U

ncer

tain

ty [m

0

0.1

0.2

0.3 PreliminaryATLAS

-1 = 13 TeV, Data 2015, 3.2 fbs

= 0.1, LCW + JES + JMST

/pcalo| < 2, mη|

= 0.2)subR = 0.05, cut

= 1.0 jets, Trimmed (fR tanti-k

caloTotal mTracking efficiencyTrack momentumFake tracks

)m

trackData/MC difference(rFragmentation modelingStatistical

(a)

[GeV]T

Jet p

210×3 310 310×2

]T

AF

ract

iona

l JM

S U

ncer

tain

ty [m

0

0.1

0.2

0.3 PreliminaryATLAS

-1 = 13 TeV, Data 2015, 3.2 fbs

= 0.1, LCW + JES + JMST

/pTA| < 2, mη|

= 0.2)subR = 0.05, cut

= 1.0 jets, Trimmed (fR tanti-k

TA mTotalTracking efficiencyTrack momentumFake tracks

)Tp

trackData/MC difference(rFragmentation modelingStatistical

calo mTotal

(b)

Figure 48: Fractional JMS uncertainty and its individual components for (a) mcalo and (b) mTA

as function of ptruthT for mreco/pT = 0.1 and |η|< 2.

10.5 C O M B I N AT I O N O F mTA A N D mcalo

The calorimeter-based jet mass is not explicitly used in deriving the track-assisted jetmass and as Figure 49 shows, the correlation between the two masses is indeed verysmall.

Arb

itrar

y un

its

0

10

20

truth / mTAm

0.5 1 1.5

trut

h /

mca

lom

0.5

1

1.5

Simulation PreliminaryATLAS

| < 2η > 250 GeV, |T

= 13 TeV, W/Z-jets, ps

correlation = 0.22

(a)

Arb

itrar

y un

its

0

2

4

6

truth / mTAm

0.5 1 1.5

trut

h /

mca

lom

0.5

1

1.5

Simulation PreliminaryATLAS

| < 2η > 1 TeV, |T

= 13 TeV, W/Z-jets, ps

correlation = 0.10

(b)

Figure 49: Correlation between the mass responses of mcalo and mTA for W/Z-jets in MC with|η|< 2 and (a) pT > 250 GeV and (b) pT > 1 TeV.

10.6 C O N C L U S I O N 77

For jets with pT > 1 TeV the correlation coefficient is just 0.1. Both mcalo and mTA areinfluenced by local fluctuations in the energy response of the calorimeter, but the im-pact on pcalo

T is much smaller. Motivated by this fact, a combination of the two masses,the so called combined jet mass, can potentially provide a more powerful mass defini-tion. Under the assumption that both individual jet mass responses are Gaussian, thecombined jet mass is given by a linear combination mcomb = a×mcalo + b×mTA. Fur-ther, since both individual jet masses are already calibrated, also mcomb is calibrated byrequiring a+ b = 1. Considering these two facts, choosing weights which are propor-tional to the individual mass resolutions is an appropriate choice for the combination:

a =σ−2

calo

σ−2calo +σ

−2TA

b =σ−2

TA

σ−2calo +σ

−2TA

(13)

In Equation 13,σcalo andσTA are the calorimeter-based and track-assisted mass resolu-tions respectively. The resolution of all three mass definitions is compared in Figure 50for W and Z-jets in MC as a function of the pT of the truth jet. At low pT mcomb ∼ mcalo

and at higher pT mcomb ∼ mTA with a smooth interpolation in the intermediate phasespace. There is a non-trivial increase in resolution of the intermediate jet pT range,and the combined JMR is always superior to the individual jet masses.

[GeV]T

Truth jet p500 1000 1500 2000 2500

)m

) / m

edia

n(R

m 6

8% IQ

nR(R

× 21

0.05

0.1

0.15

0.2

0.25


| < 0.8η = 13 TeV, W/Z-jets, |struth/mreco = mmR

calomTAmcombm

Figure 50: Resolution of mcalo, mTA and mcomb for central W/Z-jets in MC as a function of ptruthT .

The resolution is defined as half of the 68% interquantile range (IQnR) divided bythe median of the distribution.

Systematic uncertainties on mcomb are obtained by propagating the individual un-certainties on mTA and mcalo through Equation 13, while the contributions of the trackreconstruction uncertainties need to be treated as fully correlated.


The jet mass is an invaluable tool in identifying the nature of large-radius jets. Severaldifferent options exist to build this mass purely from calorimeter information, fromtracks of charged particles associated to the jet, or by combining all available informa-tion into a combined jet mass. A simulation based calibration procedure for the JMS is


inevitable, and it is demonstrated that such a technique can provide closure. At high pT

the track-assisted mass outperforms the classical definition in topologies from hadronicW /Z-boson decays in terms of both resolution and uncertainties, while providing nogain in resolution for top-jets. Combining the strengths of both the calorimeter-basedand track-assisted mass into the combined mass provides the most performant recon-structed jet mass currently available.

Part V

S E A R C H F O R R E S O N A N C E S W I T H B O S O N -TA G G E D J E T S

11S E A R C H F O R R E S O N A N C E S W I T H B O S O N -TA G G E D J E T S

The increased data set of pp collisions at the LHC atp

s = 13 TeV offers an improve-ment in sensitivity to searches for new heavy objects with masses at the TeV-scale.Diboson resonances are predicted in several extensions to the SM, such as extendedgauge symmetry models (EGM) [61–63], Grand Unified theories [64–67], theorieswith warped extra dimensions [68–72], two-Higgs-doublet models [73], little Higgsmodels [74], theories with new strong dynamics [75], including technicolour [76–78]and more generic composite Higgs models [79].

In this search, two specific benchmark models are used to optimize the event selec-tion and as a result increase the sensitivity of the search. They are also used for com-parison with the observed data. The first is the Heavy Vector Triplet Model (HVT) [80]which provides signals such as W ′→W Z and Z ′→WW , and the second a model pre-dicting a spin-2 graviton GRS decaying into WW or Z Z pairs, described by Kaluza-Kleinmodes [68, 69] of the Randall-Sundrum (RS) graviton [70–72].

With the Run 1 data ATLAS excluded at the 95% confidence level (CL) the EGMW ′→W Z with masses between 1.3 and 1.5 TeV. Upper limits were set on the produc-tion cross section times branching ratio for a GRS of the RS model decaying to WWin the fully hadronic channel. The largest excess of events over the background wasseen at 2.0 TeV in the W Z channel with a global significance of 2.5 standard devia-tions, taking the entire mass range of the search in all three channels into account [81].CMS excluded in the fully hadronic channel the EGM W ′ with masses below 1.7 TeV,and GRS with masses below 1.2 TeV, both at the 95% CL. The largest excess of eventsover the background was seen at 1.9 TeV with a global significance of 1.3 standarddeviations [82]. Upper limits on the production of generic diboson resonances usingsemileptonic final states were also published by CMS [83]. Using dileptonic final states,ATLAS has excluded at 95% CL a bulk GRS → Z Z with masses below 740 GeV [84].For narrow resonances decaying exclusively to W Z or WW , the ATLAS search in thesemileptonic channel excluded a GRS with masses below 700 GeV [85].

Both collaborations have performed searches using the early Run 2 data atp

s = 13TeV collected in 2015. In the ATLAS results [86] several final states (`νqq, ``qq andqqqq) are combined to search for narrow-width resonances with mass between 500and 3000 GeV. No significant deviations from the SM background predictions wereobserved. The data excludes at 95% CL a scalar singlet with mass below 2650 GeV,a HVT boson with mass below 2600 GeV, and a graviton with mass below 1100 GeV.Similar results from CMS [87] were presented, with no significant deviations abovethe SM predictions.

The analysis described in this chapter is a search for narrow diboson resonances(WW , W Z and Z Z) decaying into fully hadronic final states. This search is performedutilizing the combined jet mass introduced in Chapter 10, in contrast to previous re-sults from ATLAS with the same data set [88] (with significant contributions from theauthor of this thesis). W and Z bosons produced in the decay of TeV-scale resonances

81

82 S E A R C H F O R R E S O N A N C E S W I T H B O S O N -TA G G E D J E T S

are highly boosted and their decay products become increasingly collimated as themass of the resonance increases. In case they decay into quarks, they are reconstructedas a single large-radius jet. In consequence, the signature of such heavy resonance de-cays in this mode is a resonant structure in the dijet invariant mass spectrum. Thedominant multijet background produces a smoothly falling invariant mass spectrum,lacking any resonant structures. A grooming-trimming procedure is performed on thejets used in the analysis to minimize the effects of pile-up. To distinguish the signalfrom the background, jets are selected as bosonic in character, using a tagging tech-nique based on the mass of the jet and further substructure properties. This stronglysuppresses the multijet background, while still keeping a smoothly falling invariantmass distribution. The contribution to background from processes containing bosons,as V + jets (where V represents W or Z bosons), SM V V , t t̄ and single top produc-tion, is significantly smaller than from the dominant multijet background. None of thebackgrounds explored are expected to contain any resonant structures within the SM.Within the analysis, to avoid problems caused by the poorly modeled and low statis-tics backgrounds in MC samples, the expected background is modeled by a fit to thesmoothly falling distribution in data.

11.1 D ATA S A M P L E A N D S I M U L AT E D S A M P L E S

11.1.1 Data Sample

This search is performed using data collected in 2015 and 2016 fromp

s = 13 TeV LHCpp collisions with 25 ns bunch crossing separation, using the lowest-ET un-prescaledsingle large-radius jet trigger, jet ET ≥ 360 GeV in 2015 and ET ≥ 420 GeV in 2016.

For these runs the standard data quality requirements are applied to select the lumi-nosity blocks where all sub-systems of the detector were fully operational, as well aslimited periods where there was no toroidal magnetic field. The integrated luminosityafter these selections is estimated following the methodology described in [89] and is3.2 fb−1in 2015 and 12.3 fb−1in 2016. Incomplete events, as well as events flagged asbeing unusable for analysis in either the LAr or Tile calorimeters are removed.

11.1.2 Simulation of Signal and Background Events

Effects due to event pile-up are considered by overlaying additional minimum biasevents simulated with PYTHIA 8. To reproduce the pile-up conditions in data, simula-tion is reweight to match the pile-up distribution observed in data.

11.1.3 Signal Samples

The HVT and Kaluza-Klein Graviton in RS warped extra dimensions models describedbelow are used as benchmarks to optimize the analysis and ultimately set limits.

11.1 D ATA S A M P L E A N D S I M U L AT E D S A M P L E S 83

11.1.3.1 Heavy Vector Triplet Model

A generic parametrization of the couplings of a new HVT (W ′, Z ′) with the SM fieldsallows for many models to be described. The HVT phenomenological Lagrangian [80]introduces such a parametrization and allows for mixing of the SM vector bosons withthe new triplet field. Defining gV as a new parameter describing the coupling strengthto the new bosons, the coupling between the SM fermions and the new triplet can bedescribed as: g2CF /gV . Here, the gauge coupling of the SM SU(2) is denominated asg, and CF represents a multiplicative factor, modifying the coupling to SM fermions.Similarly, gV CH describes their coupling to the Higgs boson, CH again being a multi-plicative modifier.

A specific model assuming gV = 1 forming the HVT is investigated. In this scenario,the W ′ and Z ′ decay to SM particles, with a branching fraction to boson pairs (W Z ,WW , W H and ZH) of approximately 2% and a decay width of approximately 2.5%in the studied mass range. The model creates the new triplet field by an extension ofthe SM gauge group and is characterized by a weak coupling to the SM fields. Table 2lists cross sections, branching fractions and widths for selected HVT benchmarks.

Signal MC samples with masses between 1.0 and 3.0 TeV are generated using MAD-GRAPH 2.2.2 [90] interfaced to PYTHIA 8 for hadronization with the A14 tuned param-eters and the NNPDF23LO PDF sets. The samples are in intervals of 100 GeV up to2.0 TeV, then with intervals of 200 GeV up to 3.0 TeV. Only events are simulated wherethe new vector bosons decays into WW or W Z pairs which in turn decay hadroni-cally. The root mean square (RMS) of the invariant mass of the decay products of thesimulated HVT bosons is about 3.5% of the boson’s pole mass.

11.1.3.2 Kaluza-Klein Graviton in Randall-Sundrum Bulk Warped Extra Dimensions

Another model considered is the so-called bulk RS model [70], an extension to theoriginal RS model [68, 91]. Through the introduction of a warped extra dimensionit allows the SM fields to propagate in the bulk of the extra dimension. The modelhas a dimensionless coupling constant, which can be described by the curvature ofthe warped extra dimension (k) and the reduced Plank mass (MPl) as k/MPl ∼ 1 .Decays of the Kaluza–Klein graviton, GRS , into pairs of WW and Z Z bosons have aconsiderable branching fraction in this model. In the scope of this search, signals fromthe bulk RS model with k/MPl = 1 and GRS masses larger than 1 TeV are studied. Inthis model, the branching fraction of the GRS decaying to a WW (Z Z) boson pair variesbetween 18.7% and 16% (9.5% and 8%) with the graviton pole mass, and its decaywidth is approximately 2.5% in the studied mass range Table 2 lists cross sections,branching fractions and widths for selected GRS benchmarks.

Signal MC samples where the bulk GRS , with masses between 1.0 and 3.0 TeV, decaysinto WW or Z Z pairs are generated using MADGRAPH 2.2.2 interfaced to PYTHIA 8 forhadronization with the A14 tuned parameters and the NNPDF23LO PDF sets for thesame mass points as for the HVT model. The RMS of the invariant mass of the decayproducts of the simulated bulk GRS is about 10% of the its pole mass.


Table 2: Resonance width (Γ ) and cross-section times branching ratio (σ×BR) of the dibosonfinal states for different pole masses m in the described HVT and graviton model.

HVT W ′ and Z ′ GRS

WW W Z WW Z Z

m Γ σ×BR σ×BR Γ σ×BR σ×BR

[TeV] [GeV] [fb] [fb] [GeV] [fb] [fb]

1.2 40 102.1 212.7 69 29.4 14.9

1.8 56 14.4 30.6 109 2.40 1.21

2.4 74 2.94 6.37 149 0.326 0.164

3.0 92 0.726 1.58 187 0.057 0.029

11.1.4 Background Samples

The dominant background in this analysis is from the multijet QCD processes. Due tothe requirement of two high pT jets in the event, this background is well modelded bysimulated dijet processes. Due to the high background rejection of the boosted taggingcriteria and the high masses explored in this search, the dijet MC contains too fewevents in the high dijet mass tails. For the final analysis a parametrized fit is used tomodel the backgrounds from data (as described in Section 11.5). The dijet MC samplesare mainly used to test the efficiency of selection criteria, to validate the kinematicdistributions and boosted boson jet substructure variables in the background controlregions (CR), as well as to study the background fitting functions. These samples aregenerated with PYTHIA 8 with the NNPDF23LO PDF and the A14 ATLAS tune.

In addition to the dijet process, the W /Z+jets process is also considered to studythe boosted boson tagging performance. The W+jets and Z+jets events are generatedwith SHERPA 2.1.1 interfaced with the CT10 PDF set. Only the hadronic decays of theW and Z are included.

11.2 P R E S E L E C T I O N

In this section the baseline preselections used in this analysis are documented, includ-ing the trigger and physics object selections. Veto selections, which are orthogonal tothe other signal regions with similar final states, are also described.

11.2.1 Trigger Selection

Events are initially selected by the two level ATLAS trigger system. In this analysis ahigh ET, un-prescaled large-radius jet trigger1 is used. This is the lowest un-prescaledlarge-radius jet trigger present in the selected 2015 and 2016 data sets. For both datataking periods the offline selections are nearly 100% efficient with these triggers.

1 HLT_j420_a10_lcw_L1J100 for 2016 and HLT_j360_a10_lcw_sub_L1J100 for 2015.

11.3 E V E N T S E L E C T I O N 85

11.2.2 Lepton and Missing Transverse Momentum Vetoes

ATLAS performs additional searches for heavy resonant objects decaying to bosons innon-fully hadronic final states. To ensure they cover independent signal regions each ofthe searches uses an orthogonal data sample. Lepton and missing transverse momen-tum vetoes are applied. Events with central (|η| < 2.5), and high-pT (pT ≥ 25 GeV)leptons (muons or electrons) which pass both medium identification criteria [23, 92]and loose isolation requirements are rejected to place the analysis in an orthogonalregion to the (semi-)leptonic diboson channels. Events with a missing transverse mo-mentum above 250 GeV are rejected to produce orthogonality to the ννJ analysis.

11.2.3 Standard Jet Cleaning

Jets can be reconstructed from calorimeter signals that come from non-collision sourcesincluding calorimeter noise, beam halo and cosmic rays. These can be an importantbackground at high-pT, and so criteria have been developed using early 2015 data toreject such jets based on timing and jet shape. Events containing a small radius jetfailing the loose criteria [93] are rejected.

11.3 E V E N T S E L E C T I O N

11.3.1 Jet Selection

Each event must contain a pair of large-radius jets (as described in Section 6.1) whichsatisfy the following requirements. Firstly, |η| of each jet must be smaller than 2.0 toensure a good overlap with the volume of the ID, allowing the use of reconstructedtracks for boosted boson tagging (see sec. 11.3.4.1) and systematic uncertainties eval-uation [94]. After this selection, the leading-pT jet must have pT > 450 GeV, to ensurefull trigger efficiency.

The two leading jets with the highest pT passing the requirements in each event aretaken as the initial boson candidates. These are taken to arise from the decay of thesignal particle X → V V , where X is one of the theoretical particles mentioned in theintroduction and V is either a W or Z . To avoid sculpting the dijet mass spectrum bythe jet pT requirement, the invariant mass formed from the two leading jets, mJJ, isrequired to fulfill mJJ > 1.0 TeV.

11.3.2 Rapidity Difference

Signal events produced by the s-channel are more central than the t-channel QCDmultijet background events - the signal peaks near∆y12 = 0, whereas the backgroundpeaks at larger values of |∆y12|.

This selection was optimized using the modeling in MC samples for the W ′ →W Zmodels. By requiring that the dijet mass in both signal and dijet MC is within ±10 %of the generated W ′ mass, only the relevant region of phase space is selected.

Figure 51 shows the rapidity separation between the two leading trimmed anti-kt

R = 1.0 jets in QCD MC, W ′ MC and data for different dijet mass ranges corresponding

|12

y∆|E

vent

s / 0

.15

20

40

60

80

100

120310×

Data

Multi-jet MC

= 1500 GeVW'm

-1 = 13 TeV, 15.5 fbs

WZ→HVT W'

|12 |∆y0 1 2 3

Dat

a/M

C

0.5

1

1.5

(a)

|12

y∆|

Eve

nts

/ 0.1

5

2

4

6

8

10

12

310×

Data

Multi-jet MC

= 2600 GeVW'm

-1 = 13 TeV, 15.5 fbs

WZ→HVT W'

|12 |∆y0 1 2 3

Dat

a/M

C

0.5

1

1.5

(b)

Figure 51: Rapidity separation between the two leading trimmed anti-kt R = 1.0 jets in QCDMC and W ′ MC. Events in the histograms have been selected after passing triggerselection, narrow mJJ windows, |η| < 2.0 and leading jet pT > 450 GeV. Eachsubfigure displays a different dijet mass range corresponding to the relevant phasespace for the benchmark signal indicated - around (a) 1500 GeV and (b) 2600 GeV.

to the relevant phase space for the benchmark signal indicated. The ratio of signalover the square root of background (SSB) was calculated for the full range of possibleselection criteria, comparing a selection of different mass signal W ′ samples with thedominant QCD background. The optimal cut was taken as the one which provided alarge SSB ratio and signal efficiency over the full range of signal masses considered.The inputs to this optimization can be seen in Figure 52. The cut value is chosen as∆y12 < 1.2.

|12

y∆|0 0.5 1 1.5 2 2.5 3

B,S

/Bε,

Sε

0

0.5

1

1.5

2

2.5 BS/

= 1500 GeVW' mSε

Bε

Simulation = 13 TeVs

WZ→HVT W'

(a)

|12

y∆|0 0.5 1 1.5 2 2.5 3

B,S

/Bε,

Sε

0

0.5

1

1.5

2

2.5 BS/

= 2600 GeVW' mSε

Bε


WZ→HVT W'

(b)

Figure 52: Optimization of the rapidity separation selection between the two leading trimmedanti-kt R = 1.0 jets. Events in the histograms have been selected after passingtrigger selection, narrow mJJ windows, |η| < 2.0, leading jet pT > 450 GeV. Eachsubfigure displays a different dijet mass range corresponding to the relevant phasespace for the benchmark signal indicated - around (a) 1500 GeV and (b) 2600 GeV.The SSB (green) for a range of selection criteria are shown for a sub-set of W ′

signal masses. In red and in blue are shown respectively the signal and backgroundefficiency for different value of rapidity separation cut.

11.3.3 pT Asymmetry

Only two highly energetic jets are expected as the final state for the considered signalmodels. These two jets should form a balanced system, so there pT should be nearlyidentical. This balance can be quantified in form of the pT asymmetry defined as

A=pT1− pT2

pT1 + pT2, (14)

where pT1 and pT2 are the pT of the leading and sub-leading jets. Without any radiationin the final or initial states and the lack of detector or reconstruction effects whichresult in a jet pT resolution different from zero, Awould be zero. This is true for both theconsidered signals and the dijet background. This cut is therefore applied to removeany event where either of the jets is badly reconstructed, rather than to suppress thebackground. Figure 53 shows the distribution of the pT asymmetry in data and signaland background MC events.

The selection requirement placed on this variable was optimized using the modelingin MC samples, applying |∆y12|< 1.2 and the same phase space and baseline selectionused for the optimization of the rapidity separation cut. The SSB was calculated forthe full range of possible selection criteria, comparing a selection of different masssignal W ′ samples to the dominant QCD background. The final cut value A< 0.15 waschosen to still be on the plateau of the selection efficiency for the signal. The inputs tothis optimization can be seen in Figure 54.

11.3.4 Boosted Boson Tagging

To enhance the separation between the signal boson jets and the QCD jets, several dis-criminating jet substructure variables have been studied in the past [94]. The largestseparation between signal and background jets has been achieved by applying addi-tional cuts on three discriminating variables:

Trimmed jet mass: the trimmed jet mass should be consistent with the boson massfor jets from hadronically decaying bosons while for jets from the background itshould be small .

D2: boson jets are characterized by a two-prong structure, while QCD jets are mainlyone-prong. This discriminant is a variation on the ratio of energy correlationswhich optimizes the separation between one-prong and two-prong decays, inanalytical terms [94].

Number of tracks associated to the untrimmed jet, ntrk: background events containinghard gluons are likely to pass the D2 cut. Using the fact that energetic gluonemissions, on average, produce jets with high charged particle multiplicities, thiscut can remove them.

Figure 55 shows a diboson candidate event passing the full analysis selection. Thetwo-prong structure one expects for these jets, and to which the D2 selection is sensi-tive, can be observed for both jets.

|12

y∆|

Eve

nts

/ 0.0

1

10

20

30

40

50

60

70

80

90

310×

Data

Multi-jet MC

= 1500 GeVW'm

-1 = 13 TeV, 15.5 fbs

WZ→HVT W'

AsymmetryT

p0 0.1 0.2 0.3

Dat

a/M

C

0.5

1

1.5

(a)

|12

y∆|

Eve

nts

/ 0.0

1

200

400

600

800

1000

1200

1400

1600

1800 Data

Multi-jet MC

= 2600 GeVW'm

-1 = 13 TeV, 15.5 fbs

WZ→HVT W'

AsymmetryT

p0 0.1 0.2 0.3

Dat

a/M

C

0.5

1

1.5

(b)

Figure 53: Transverse momentum asymmetry distribution between leading and sub-leadingtrimmed anti-kt R = 1.0 jets in QCD MC and W ′ MC. Events in the histogramshave been selected after passing trigger selection, narrow mJJ windows, |η| < 2.0,leading jet pT > 450 GeV and |∆y| < 1.2. Each subfigure displays a different dijetmass range corresponding to the relevant phase space for the benchmark signalindicated - around (a) 1500 GeV and (b) 2600 GeV.

AsymmetryT

p

0 0.05 0.1 0.15 0.2 0.25

B,S

/Bε,

Sε

0

0.2

0.4

0.6

0.8

1

1.2

1.4 BS/

= 1500 GeVW' mSε

Bε


WZ→HVT W'

(a)

AsymmetryT

p

0 0.05 0.1 0.15 0.2 0.25

B,S

/Bε,

Sε

0

0.2

0.4

0.6

0.8

1

1.2

1.4 BS/

= 2600 GeVW' mSε

Bε


WZ→HVT W'

(b)

Figure 54: Optimization of the transverse momentum asymmetry selection between leadingand sub-leading trimmed anti-kt R = 1.0 jets. The QCD and signal MC are normal-ized to unity. Events in the histograms have been selected after a cut of |η|< 2.0, a|∆y| < 1.2 requirement and the trigger plateau is modeled by a pT > 450 GeV re-quirement on the input jets. Each subfigure displays a different dijet mass range cor-responding to the relevant phase space for the benchmark signal indicated - around(a) 1500 GeV and (b) 2600 GeV. The SSB (green) for a range of selection criteriaare shown for a sub-set of W ′ signal masses. The signal and background efficiencyfor different values of pT balance are shown in red and blue.


Figure 55: Event display of a diboson candidate event recorded by the ATLAS detector on 20thof May 2016. The dijet invariant mass is 2.4 TeV. The leading (sub-leading) jettransverse momenta is 1.26 TeV (1.15 TeV). The leading (sub-leading) jet invariantmass is 83.7 GeV (73.9 GeV). The ntrk (D2) is 25 (1.62) for leading jet and 24 (1.34)for sub-leading jet.

The optimization and definition of jet mass and D2 cuts are described in Refer-ence [94]. Boosted boson tagging based on trimmed anti-kt R = 1.0 jets with cuts onjet mass and D2 is currently used in all ATLAS analyses which include hadronic bosonsin the final state. A working point with a constant 50 % boson selection efficiency hasbeen chosen for this analysis. The selection on the number of tracks associated withthe untrimmed jet increase the QCD rejection power, while preserving high signal ef-ficiency, and it has been designed and optimized specifically for this analysis.

11.3.4.1 Track Multiplicity

Both mass and D2 cuts can be passed by background jets containing radiation froma hard gluon, because they are likely to exhibit a two prong structure as expectedfrom the signal. With increasing energy, the number of charged hadrons increasesfor both gluon and quark induced jets [95]. Even so, on average, gluon-induced jetsproduce a much higher charged particle multiplicity than quark-induced jets of thesame energy. The relevant scale for QCD dijet background is the momentum of thejet, while for decays of boosted bosons it is the boson mass. Since the former is muchlower, one can use this to discriminate between the two processes. By cutting on themultiplicity of tracks associated to the trimmed jet, ntrk (proportional to the chargedhadron multiplicity), one can increase the background rejection while retaining mostof the signal efficiency.

Tracks fulfilling the minimum ID quality requirements and matched to the eventprimary vertex, are associated to jets using the ghost association procedure. Figures 56and 57 show the distribution of the number of tracks associated with large-radius jetsbefore trimming, for HVT W ′ signals, dijet MC and data passing the same selections

|12

y∆|E

vent

s / 1

.00

100

200

300

400

500

600

700Data

Multi-jet MC

= 1500 GeVW'm

-1 = 13 TeV, 15.5 fbs

WZ→HVT W'

trk

0 20 40 60 Leading jet n

Dat

a/M

C

0.5

1

1.5

(a)

|12

y∆|

Eve

nts

/ 1.0

0

20

40

60

80

100 Data

Multi-jet MC

= 2200 GeVW'm

-1 = 13 TeV, 15.5 fbs

WZ→HVT W'

trk

0 20 40 60 Leading jet n

Dat

a/M

C

0.5

1

1.5

(b)

Figure 56: Transverse trimmed anti-kt R = 1.0 jets in QCD MC and W ′ MC. ntrk for the leadingjet. The QCD and signal MC are normalized to data. Events in the histograms havebeen selected after passing trigger selection, |η| < 2.0, leading jet pT > 450 GeV,|∆y|< 1.2, A< 0.15 as well as partial boson tagging. Two different dijet mass rangecorresponding to the relevant phase space for the benchmark signal indicated areshown - around (a) 1500 GeV and (b) 2200 GeV.

as for the optimization described in the following. The selection requirement placedon this variable is optimized using the modeling in MC samples, requiring that all theevents have passed the baseline selections as well as containing two large-radius jetssuccessfully tagged as a boson by the D2 and jet mass requirements. The optimal cut,ntrk < 30, was taken as the one which provided a large SSB ratio and signal efficiencyover the full range of signal masses considered, as seen in Figure 58. Using ntrk toidentify weak boson jets improves the expected sensitivity by approximately 20–30%.

|12

y∆|

Eve

nts

/ 1.0

0

100

200

300

400

500Data

Multi-jet MC

= 1500 GeVW'm

-1 = 13 TeV, 15.5 fbs

WZ→HVT W'

trk

0 20 40 60 Sub-leading jet n

Dat

a/M

C

0.5

1

1.5

(a)

|12

y∆|

Eve

nts

/ 1.0

0

10

20

30

40

50

60

70

80Data

Multi-jet MC

= 2200 GeVW'm

-1 = 13 TeV, 15.5 fbs

WZ→HVT W'

trk

0 20 40 60 Sub-leading jet n

Dat

a/M

C

0.5

1

1.5

(b)

Figure 57: Transverse trimmed anti-kt R = 1.0 jets in QCD MC and W ′ MC. ntrk for the sub-leading jet. The QCD and signal MC are normalized to data. Events in the his-tograms have been selected after passing trigger selection, |η| < 2.0, leading jetpT > 450 GeV, |∆y| < 1.2, A < 0.15 as well as partial boson tagging. Two differ-ent dijet mass range corresponding to the relevant phase space for the benchmarksignal indicated are shown - around (a) 1500 GeV and (b) 2200 GeV.

trkn

0 5 10 15 20 25 30 35 40 45

B,S

/Bε,

Sε

0

0.5

1

1.5

2

2.5BS/

= 1500 GeVW' mSε

Bε


WZ→HVT W'

(a)

trkn

0 5 10 15 20 25 30 35 40 45

B,S

/Bε,

Sε

0

0.5

1

1.5

2

2.5BS/

= 2200 GeVW' mSε

Bε


WZ→HVT W'

(b)

Figure 58: Optimization of the ntrk selection. Events in the histograms have been selected afterpassing trigger selection, narrow mJJ windows, |η|< 2.0, leading jet pT > 450 GeV,|∆y| < 1.2 and A < 0.15 as well as partial boson tagging. The optimization plotsare showed for (a) 1.5 and (b) 2.2 TeV W ′ signal samples, corresponding to therelevant phase space for the benchmark signal indicated. The SSB is in green, whilein red and in blue are shown respectively the signal and background efficiency fordifferent value of ntrk cut.

11.4 C O N T R O L R E G I O N S 95

The pile-up dependence of ntrk has been studied and found to be negligible. A datasample of V+jets events is used to estimate the relative efficiency and backgroundrejection of the ntrk cut, as described in Section 11.4.3.

Figure 59 shows the selection efficiency for the HVT and bulk GRS signals as a func-tion of the resonance mass. For the WW and W Z decay in the HVT model, and theGRS → WW decay in the bulk RS model, the acceptance times selection efficiency isaround 0.12. The GRS → Z Z decay in the bulk RS model has a lower acceptance timesselection efficiency around 0.09 to 0.1.

1500 2000 2500 3000m [GeV]

Acc

×Ef

f

0

0.05

0.1

0.15

0.2

0.25

0.3

= 13 TeVsHVT W'→WZHVT Z'→WWG →WWRS

→ZZRSG

Figure 59: Selection efficiency for HVT model Z ′→WW , GRS →WW , HVT model W ′→W Zand GRS → Z Z benchmark signals after applying either the full WW , W Z or Z Zselections.

11.4 C O N T R O L R E G I O N S

Various data CR are used to evaluate the efficiency of selections made in the analy-sis or the effectiveness and flexibility of the background modeling procedure. A briefdescription of each follows here.

11.4.1 Mass Sideband Regions

Alternative mass window selections to the boson tagger are used to test the effective-ness of the background modeling function without unblinding the signal region. Thefollowing sidebands are considered:

• A low mass sideband region: Both leading jets are required to have 50 < mJ <

65.

• A high mass sideband region: Both leading jets are required to have 110< mJ <

140. However, this mass window contains the signal region of analyses usingHiggs in the event. To avoid the possibly unblinding of the signal region of e.g.vector-associated Higgs production a veto on b-tagged large-R jets is imposed. A

|12

y∆|

Eve

nts

/ 100

.00

1−10

1

10

210

310

410

510

610Data

Multi-jet MC

s = 13 TeV, 15.5 fb-1

low-low CR

1000 1500 2000 2500 3000 3500

Dat

a/M

C

0.5

1

1.5

mJJ [GeV]

(a)

|12

y∆|

Eve

nts

/ 100

.00

1−10

1

10

210

310

410

510

610Data

Multi-jet MC

s = 13 TeV, 15.5 fb-1

low-high CR

1000 1500 2000 2500 3000 3500

Dat

a/M

C

0.5

1

1.5

mJJ [GeV]

(b)

|12

y∆|

Eve

nts

/ 100

.00

1−10

1

10

210

310

410

510

610Data

Multi-jet MC

s = 13 TeV, 15.5 fb-1

high-low CR

1000 1500 2000 2500 3000 3500

Dat

a/M

C

0.5

1

1.5

mJJ [GeV]

(c)

|12

y∆|E

vent

s / 1

00.0

01−10

1

10

210

310

410

510

610

Data

Multi-jet MC

s = 13 TeV, 15.5 fb-1

high-high CR

1000 1500 2000 2500 3000 3500

Dat

a/M

C

0.5

1

1.5

mJJ [GeV]

(d)

Figure 60: Invariant dijet mass distributions in data and MC simulation for the (a) low-lowand (b) low-high (c) high-low and (d) high-high sideband CR.

large-R jet is b-tagged if more than one of the ghost-associated R = 0.2 subjetsis b-tagged.

• Mixed mass sideband regions: Using combinations of low/high mass windowcuts on the leading and sub-leading jets to form additional sideband regions.

The invariant dijet mass spectrum for all four sideband CR in data and MC are shown inFigure 60. Agreement within 10% can be observed for mJJ < 1.5 TeV where statisticaluncertainties do not dominate.

11.4.2 Partial-tag Control Regions

The performance of the boson tagging requirements and their impact on the invariantdijet mass spectrum is important to understand if the jet mass, D2 or ntrk requirementsare sculpting the background. The mJJ distribution after applying either a selectionon the jet mass window, the energy correlation variable D2 or the ntrk are shown inFig. 61. No sculpting of the distribution in either MC or data can be observed. Eventhough large statistical fluctuations can be observed above approximately 1.5 TeV, thelow mJJ spectra in data is well reproduced by MC.

11.4 C O N T R O L R E G I O N S 97

|12

y∆|

Eve

nts

/ 100

.00

1−10

1

10

210

310

410

510

610 Data

Multi-jet MC

s = 13 TeV, 15.5 fb-1

partial tag: jet mass

1000 1500 2000 2500 3000 3500

Dat

a/M

C

0.5

1

1.5

mJJ [GeV]

(a)

|12

y∆|

Eve

nts

/ 100

.00

1−10

1

10

210

310

410

510

610 Data

Multi-jet MC

s = 13 TeV, 15.5 fb-1

2partial tag: D

1000 1500 2000 2500 3000 3500

Dat

a/M

C

0.5

1

1.5

mJJ [GeV]

(b)

|12

y∆|

Eve

nts

/ 100

.00

1−10

1

10

210

310

410

510

610

710

Data

Multi-jet MC

s = 13 TeV, 15.5 fb-1

trkpartial tag: n

1000 1500 2000 2500 3000 3500

Dat

a/M

C

0.5

1

1.5

mJJ [GeV]

(c)

Figure 61: Invariant dijet mass distributions in data and MC simulation after requiring (a) theleading and sub-leading jet to be within the mass window of the boson tagger, (b)after applying the D2 criteria derived for a 50% working point and (c) after applyingthe ntrk selection.

11.4.3 V+jets CR

A sample enriched in vector boson plus jets is obtained starting from the pre-selectionof two jets in an |η|<2.0 region. Further, it is requested that the leading jet has a pT

larger than 550 GeV and successfully is tagged by the D2 selection as either a W or aZ . The resulting events have a small peak in the mass distribution which can be usedto fit the rate of W + Z signal, even in bins of the number of charged tracks. Using thedata/MC ratio of the fitted rate of W/Z bosons as a function of ntrk the quality of thedescription of the ntrk variable in MC simulation can be assessed.

The mJ distribution in the CR for ntrk < 30 is shown in Figure 62 (a) - black dots rep-resent data and red squares denote simulation. The fitted mJ distribution, composedof the multijet background and the W and Z peaks, is plotted as a solid black line, andthe dashed red line shows solely the background component. Both mass and width ofthe fitted W + Z peak in data are well modeled by the MC simulation. Figure 62 (b)shows the fitted fraction of W /Z jets for data and simulation as a function of ntrk.Two distributions are shown for the simulated ntrk, one without modifications andone where the distribution is scaled by 1.06 to best match the distribution observed indata. This scale difference is used as a systematic uncertainty on the ntrk modeling.

11.5 B A C K G R O U N D PA R A M E T R I Z AT I O N

The search for diboson resonances is performed by looking for narrow peaks above thesmoothly falling mJJ distribution expected in the SM. This smoothly falling backgroundis overwhelmingly composed of SM QCD dijet events. Other processes, which also fallsmoothly as a function of mJJ, like SM dibosons, W /Z + jets and t t, have small (belowfew percent) to negligible contributions to the background. The background to thesearch is estimated empirically from the observed mJJ spectrum in the signal region.The background estimation procedure is based on a binned maximum-likelihood fit ofthe observed mJJ spectrum to a parametric form. This is given by

dnd x

= p1(1− x)p2+ξp3 x p3 (15)

where x = mJJ/p

s, p1 is a normalization factor, p2 and p3 are dimensionless shapeparameters, and ξ is a constant chosen to minimize the correlation between p2 andp3 in the fit. The fit is performed on a histogram of the observed mJJ distribution indata with a constant bin size of 100 GeV. The fit range is 1.0 < mJJ < 3.5 TeV, wherethe lower bound is dictated by the point where the trigger is fully efficient for boson-tagged jets. The upper bound is set because the D2 parametrization is available forjets with pT only up to 2 TeV and therefore the signal selection efficiency drops after3.5 TeV affecting the fit.

The modeling of the parametric shape in Equation 15 is tested in simulation andthe background enriched CRs in data defined in Section 11.4. Figure 63 shows the fitresults performed in the high and mixed mass CRs in data and Table 3 list the resultsof the fits.

11.5 B A C K G R O U N D PA R A M E T R I Z AT I O N 99

Eve

nts

/ 5 G

eV

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

trk

s=13 TeV, 15.5 fb-1

n <30

ATLAS Preliminary

DataMulti-jet MCW+jets MCZ+jets MCFitted s+bFitted bkd.

[GeV]Jm40 60 80 100 120 140 160 180 200 220 240

Dat

a-fit

0500

1000 450±Signal=6599

(a)

trk n0 10 20 30 40 50 60 70 80 90

Fra

ctio

n of

W/Z

eve

nts

/ tra

ck

0

0.01

0.02

0.03

0.04

0.05

Data

W/Z MC

W/Z reweighted

ATLAS Preliminary

-1=13 TeV, 15.5 fbs

(b)

Figure 62: (a) Comparison of the mJ distribution between data and dijet simulation for jetswith pjet

T > 550 GeV and passing the D2 boson requirement. The expected shapeof the boson peak from W/Z+jets simulation is also shown. On the bottom, thedifference between data and the fitted background is shown. (b) Comparison ofrelative fraction of W /Z-jets as a function of ntrk. Simulation is shown with andwithout a reweighting based on the fitted scale difference to data. Statistical errorsare indicated by black error bars for data. The red error bars include systematicuncertainties of the method.


1500 2000 2500 3000 3500E

vent

s / 1

00 G

eV

1

10

210

310

410-1 = 13 TeV, 15.5 fbs

high-high CRDataFit bkg estimation

mJJ [GeV]1500 2000 2500 3000 3500

Pul

l

2−

0

2

(a)

1500 2000 2500 3000 3500

Eve

nts

/ 100

GeV

1

10

210

310

410

-1 = 13 TeV, 15.5 fbslow-high CR

DataFit bkg estimation

m JJ [GeV]1500 2000 2500 3000 3500

Pul

l

2−

0

2

(b)

Figure 63: Comparison between the fitted mJJ shape (statistical uncertainty shown as grayerror band) and the mJJ data spectra in CRs. In (a) events are selected requiringto be in the high-high sideband mass CR and in (b) to be in the low-high sidebandmass CR. The bottom plot in the figures show the pull, defined as the z-value asdescribed in Reference [96].

Table 3: Goodness-of-fit for maximum-likelihood fits of the background model to the dijet massdistribution in high, low and mix mass CR.

Sample χ2/nDOF p-value

Data in low-low CR 19/22 0.63

Data in high-high CR 7.0/21 1.00

Data in low-high CR 20.8/22 0.54

11.6 S Y S T E M AT I C U N C E RTA I N T I E S 101

11.6 S Y S T E M AT I C U N C E RTA I N T I E S

The uncertainties effecting the background modeling are taken directly from the er-rors on the fit parameters of the background estimation procedure described in Sec-tion 11.5. The systematic uncertainties on the expected signal yield and shapes arisefrom detector effects and MC modeling. These are evaluated using the signal MC sam-ples over the search region 1.0 < mX < 3.5 TeV. Systematic uncertainties on signalshape and normalization are assessed and expressed in terms of nuisance parametersin the statistical analysis as shown in Section 11.7.2. Uncertainties due to the large-radius jet calibrations affect the large-radius jet pT, mass and D2. It is assumed thatpT and mass scale are fully correlated, while the scale of D2 is uncorrelated with theformer two.

11.6.1 Jet Energy and Mass Scale

Uncertainties on the JES are an important effect in the search for resonant structuresin the presence of a rapidly falling background spectrum. This uncertainty shifts theexpected signal mass spectrum, particularly the peak of the resonance, affecting thesignificance of an excess if observed. Uncertainty in the JMS affects the observed jetmass, which in turn affects the boosted tag selection efficiency and the dijet massshape.

The JES and JMS systematic uncertainties on the trimmed large-radius jets are eval-uated using the methods described in Section 10.4. The size of the total JES and JMSuncertainties vary with jet pT and are around 5% and 3% for the full mass range.

11.6.2 Jet Energy and Mass Resolution

Uncertainties in the measurement of the jet energy resolution (JER) and JMR wouldlead to a mis-measurement of the width of any observed signal and affect the signalselection efficiency. An uncertainty in the measurement of the JMR smears out the ob-served jet mass distribution. The JER and JMR uncertainties are evaluated by applyingan additional relative Gaussian smearing on the input jets energy and mass, degradingthe nominal resolution by 20%. This is performed in four bins of jet pT, 250−500 GeV,500−1000 GeV, 1.0−1.5 TeV and ≥ 1.5 TeV to account for the dependence of jet pT.

11.6.3 D2 Scale and Resolution

Any uncertainty on the value of the boson tagging discriminant D2 used here wouldaffect the selection efficiency of the analysis. Any variation observed will be taken asa systematic uncertainty on the value of the cut, evaluated by shifting the cut valueby the uncertainty. The systematic uncertainty on the trimmed large-radius jets usedin this analysis has been evaluated using the methods described in Section 10.4. Thesize of the uncertainty varies with jet pT and is around 6% for the full mass range.

Uncertainty in the D2 resolution acts to smear out the D2 distribution and thus al-ters the efficiency of the final analysis cut. This is evaluated as an additional relativeGaussian smearing of the input jets D2, degrading the nominal resolution by 10%. This


is performed in four bins of jet pT, 250− 500 GeV, 500− 1000 GeV and ≥ 1.0 TeV asthe nominal D2 resolution varies with jet pT. The response of the D2 cut is not strictlyGaussian and thus the RMS of the observed distribution is taken as a conservativeapproximation to the width.

11.6.4 Track-multiplicity Efficiency

Uncertainties on the number of tracks associated with the untrimmed jet can affect thesignal selection efficiency. The modeling of this variable is discussed in Section 11.4.3,and an uncertainty of 6% is estimated.

11.6.5 Luminosity Scale

The preliminary uncertainty on the integrated luminosity of both the 2015 and 2016data sets is 3.0%. It is derived from a preliminary calibration of the luminosity scaleusing x-y beam separation scans performed in August 2015 and May 2016. A detaileddescription of this method is given in Reference [89] and [97].

11.7 R E S U LT S

11.7.1 Background Only Fit Results in Signal Regions

A background-only binned maximum likelihood fit is first performed on the observeddijet invariant mass distribution in data, with the background modeling function de-scribed in Section 11.5. Figure 64 shows the dijet invariant mass distribution observedin the W Z , WW and Z Z signal region in data, compared to the fitted backgroundshape. The uncertainties of the fitted parameters are propagated to the dijet mass dis-tribution and are shown as shaded-bands. The lower panels in the figure show thesignificances, defined as the signed z-value of the difference between the observeddata and the expectation [96]. The fitted values of the parameters are summarized inTable 4. No significant excess is observed in data.

Table 4: Fitted parameters of the dijet mass distribution and number of observed events foreach signal region.

Parameter WW W Z Z Z

ξ 7.54 7.79 7.72

p2 51.0±2.0 53.3±1.8 54.6±2.3

p3 7.8±1.4 6.6±1.2 7.9±1.6

Observed events 491 709 421

11.7 R E S U LT S 103

1000 1500 2000 2500 3000 3500

Eve

nts

/ 100

GeV

1−10

1

10

210

310

410-1 = 13 TeV, 15.5 fbs

WW selectionDataFit bkg estimation

[GeV]JJm1000 1500 2000 2500 3000 3500

Pul

l

2−

0

2

(a)

1000 1500 2000 2500 3000 3500

Eve

nts

/ 100

GeV

1

10

210

310-1 = 13 TeV, 15.5 fbs

WZ selectionDataFit bkg estimation

[GeV]JJm1000 1500 2000 2500 3000 3500

Pul

l

2−

0

2

(b)

1000 1500 2000 2500 3000 3500

Eve

nts

/ 100

GeV

1

10

210

310-1 = 13 TeV, 15.5 fbs

ZZ selectionDataFit bkg estimation

[GeV]JJm1000 1500 2000 2500 3000 3500

Pul

l

2−

0

2

(c)

Figure 64: The observed distributions in data in the (a) WW , (b) W Z and (c) Z Z signal regions.The fitted background is also shown, where the shaded-bands are its uncertainty.


11.7.2 Statistical Analysis

The data is interpreted with a frequentist analysis by calculating the signal strength, µfor each of the benchmark models. µ scales with the number of signal events predictedby the analysis, and for the signal plus background hypothesis equals unity. In contrast,µ is zero for the background only hypothesis. A test statistic, λ(µ), based on the profilelikelihood ratio [98] is used to test the two benchmark models. A maximum likelihoodfit of the signal plus background hypothesis to the data extracts the value of µ. If anyexcess is observed over the background only hypothesis, it is quantified in terms of localp0. It corresponds to the probability of the background only hypothesis to producean excess at least as large as the one observed. The largest p0 in this search has alocal significance of 2.1 standard deviations for a HVT W ′ and a RS graviton decayingto WW at a resonance mass of 1.2 TeV as shown in Figure 65. Such a deviation isconsistent with the expected background fluctuations.

[GeV]JJm1500 2000 2500 3000

0Lo

cal p

8−10

5−10

2−10

10

410

610-1 = 13 TeV, 15.5 fbs

σ0σ1σ2

σ3

σ4

σ5

WZ→HVT W’

ZZ→ RSG

WW→HVT Z’

WW→ RSG

Figure 65: Observed p0-values for the HVT Z ′ →WW , and W ′ →W Z , and bulk RS gravitonGRS →WW and GRS → Z Z signal models. Very similar p0-values are expected forboth WW signals, since they both produce similar signal shapes and the used datadistribution is identical.

Limits on the production cross section times branching ratio of each consideredsignal model are set as a function of the resonance mass. For the W Z selections, theHVT W ′ is used as benchmark and for the Z Z selection, the bulk GRS is used. Both Z ′

and GRS are used as benchmark for the WW selection. Figure 66 (a) and Figure 66 (b),show the observed 95% CL upper limits on the cross section times branching ratio onthe HVT W ′ → W Z and Z ′ → WW hypotheses as function of the W ′/Z ′ mass. Newresonances of the HVT model are excluded for masses between 1.2-2.0 TeV for the WZchannel, and between 1.3-1.7 TeV for the WW channel with 95% CL.

Figures 66 (c) and Figure 66 (d) show the observed 95% CL upper limits on thecross section times branching ratio for the bulk GRS → Z Z and WW respectively. Noexclusion can be made for this model, since the cross section times branching ratio forexcited graviton production is below the sensitivity of this analysis.

11.7 R E S U LT S 105

[GeV]Z'm1500 2000 2500 3000

WW

) [fb

]→

BR

(Z'

×Z

'+X

)→

(pp

σ

1−10

1

10

210

310

410

-1 = 13 TeV, 15.5 fbsObserved 95% CLExpected 95% CL

σ 1±σ 2±

WW→HVT Z'

(a)

[GeV]W'm1500 2000 2500 3000

WZ

) [fb

]→

BR

(W'

×W

'+X

)→

(pp

σ

1−10

1

10

210

310

410


σ 1±σ 2±

WZ→HVT W'

(b)

[GeV]RS

Gm1500 2000 2500 3000

WW

) [fb

]→

RS

BR

(G×

+X

)R

SG

→(p

pσ

1−10

1

10

210

310

410


σ 1±σ 2±

=1Pl

MWW, k/→RSG

(c)

[GeV]RS

Gm1500 2000 2500 3000

ZZ

) [fb

]→

RS

BR

(G×

+X

)R

SG

→(p

pσ

1−10

1

10

210

310

410


σ 1±σ 2±

=1Pl

MZZ, k/→RSG

(d)

Figure 66: Observed and expected 95% CL limits on the cross-section times branching ratiofor (a) HVT Z ′→WW , (b) HVT W ′→W Z , (c) bulk RS graviton GRS →WW and(d) bulk RS graviton GRS → Z Z channels to diboson final states.

12C O N C L U S I O N

The Large Hadron Collider entered a new energy regime at Run 2 with proton–protoncollisions at

ps = 13 TeV. Events with multi-TeV jets showering in the detectors, or

tau-leptons and b-hadrons surviving passage through multiple active layers of mate-rial, are now common place. These objects are also signatures for new physics, in-cluding massive particles that decay to highly boosted bosons, whose own subsequentdecay products are often reconstructed into one large-radius jet. One of the strongestdiscriminants between these jets and the multijet background is their mass.

In the core of highly energetic hadronic jets, the average separation of charged par-ticles is comparable to the size of individual inner detector elements. As a result, theircharge deposits in the pixel and SCT detector start to overlap and can be reconstructedas a single merged cluster. This can create confusion within the algorithms reconstruct-ing charged particle trajectories (tracks). Without careful consideration, the track re-construction efficiency in these dense environments will be limited.

An overview of the offline track reconstruction in ATLAS is given, and the workleading to a vastly improved setup in terms of performance in dense environments forRun 2 is presented. Key developments in the ambiguity solver, in combination witha novel use of a neural network to identify merged pixel clusters, are presented. Theimproved performance is demonstrated for highly collimated tracks from decays ofsingle particles, as well as in the more physical environment of hadronic Z ′ decays inthe presence of event pile-up. Up to 10% more pixel clusters are associated to tracksin the core of high pT jets, which results in a much more robust track reconstructionefficiency. This improvement becomes most apparent for charged particles with a highproduction radius (>30 mm), where 17% efficiency is recovered, as well as in thecore of high pT b- and light-jets, where 10% and 14% efficiency is recovered. Thehigher quality and efficiency of reconstructed tracks directly boosts the performanceof several derived physics objects, which is demonstrated with the example of flavortagging. For flavor tagging a 7–13% improvement in b-jet efficiency is achieved for afixed light-jet rejection for jets with pT > 100 GeV, using an IP3D tagger which has notbeen re-optimized.

Extending these Monte Carlo based studies, two methods are introduced to probethe track reconstruction performance in data. First, a study of the properties of pixelclusters for collimated track pairs allows for a comparison of the identification effi-ciency of merged pixel clusters in data and simulation. For separations between thepair of tracks below the dimension of a single pixel, the efficiency for identifying thecluster as merged and assigning it correctly to both tracks is above 80% for the IBLand above 90% for the B-layer. The results are consistent between data and simulation,with small residual discrepancies. Using data alone, the second method is able to quan-tify a residual inefficiency of the track reconstruction in the core of jets as a functionof the transverse momentum of the jet using the energy loss in silicon. It varies from0.061± 0.006(stat.)± 0.014(syst.) to 0.093± 0.017(stat.)± 0.021(syst.) between atransverse jet momentum of 200 to 400 GeV and 1400 to 1600 GeV, respectively.

107

108 C O N C L U S I O N

With this improved track reconstruction performance, and vastly smaller uncer-tainties through the data-driven measurement of the track reconstruction inefficiency,ways to reconstruct the masses of jets with higher precision become possible. Threedifferent options exist to build the mass of the jet purely from calorimeter information,from tracks of charged particles associated to the jet, or by combining all available in-formation into a combined jet mass. A simulation based calibration procedure for thejet mass scale is inevitable, and it is demonstrated that such a technique can provideclosure. Combining the strengths of both the calorimeter and tracker into the com-bined mass provides the most performant reconstructed jet mass currently available,reducing both the resolution of the reconstructed mass of the jet and its uncertainty.

Employing these novel reconstruction methods, a search for resonances with massesin the range 1.2 < m < 3.5 TeV in the hadronically decaying W Z , WW , or Z Z finalstate, is performed in 15.5 fb−1 of

ps = 13 TeV proton-proton collision data. No signifi-

cant deviations from the background expectations are observed. An additional chargedor neutral heavy vector boson, as predicted by the Heavy Vector Triplet phenomeno-logical Lagrangian (assuming gV = 1), decaying through W ′ → W Z (or Z ′ → WW ),is excluded in the mass range 1.2–2.0 (1.3–1.7) TeV at the 95% confidence level.

B I B L I O G R A P H Y

[1] A. Einstein. Die Grundlage der allgemeinen Relativitätstheorie. Ann. Phys.,354:769–822, 1916.

[2] C. Patrignani and Particle Data Group. Review of particle physics. Chin. Phys. C,40(10):100001, 2016.

[3] M. Gell-Mann. A schematic model of baryons and mesons. Phys. Lett., 8:214–215,1964.

[4] S. L. Glashow. Partial symmetries of weak interactions. Nucl. Phys., 22:579–588,1961.

[5] S. Weinberg. A model of leptons. Phys. Rev. Lett., 19:1264–1266, 1967.

[6] A. Salam. Gauge unification of fundamental forces. Rev. Mod. Phys., 52:525–538,Jul 1980.

[7] F. Englert and R. Brout. Broken symmetry and the mass of gauge vector mesons.Phys. Rev. Lett., 13:321–323, 1964.

[8] P. W. Higgs. Broken symmetries and the masses of gauge bosons. Phys. Rev. Lett.,13:508–509, 1964.

[9] UA1 Collaboration. Experimental observation of isolated large transverse energyelectrons with associated missing energy at

ps= 540 GeV. Phys. Lett., B122:103–

116, 1983.

[10] UA1 Collaboration. Experimental observation of lepton pairs of invariant massaround 95 GeV/c2 at the CERN SPS collider. Phys. Lett., B126:398–410, 1983.

[11] ATLAS Collaboration. Observation of a new particle in the search for the Stan-dard Model Higgs boson with the ATLAS detector at the LHC. Phys. Lett. B,716:1–29, 2012.

[12] CMS Collaboration. Observation of a new boson at a mass of 125 GeV with theCMS experiment at the LHC. Phys. Lett. B, 716:30–61, 2012.

[13] ATLAS Collaboration. The ATLAS experiment at the CERN Large Hadron Collider.JINST, 3(08):S08003, 2008.

[14] L. Evans and P. Bryant. LHC machine. JINST, 3(08):S08001, 2008.

[15] K. Sliwa. ATLAS overview and main results. In Proceedings, International Schoolon High Energy Physics : Workshop on High Energy Physics in the near Future.(LISHEP 2013): Rio de Janeiro, Brazil, March 17-24, 2013, 2013.

[16] ATLAS Collaboration. ATLAS Inner Detector: Technical Design Report. 1997.

109

110 B I B L I O G R A P H Y

[17] ATLAS Collaboration. ATLAS pixel detector electronics and sensors. JINST,3(07):P07007, 2008.

[18] R. W. L. Jones and D. Barberis. The evolution of the ATLAS computing model. J.Phys. Conf. Ser., 219(7):072037, 2010.

[19] ATLAS Collaboration. ATLAS computing: Technical Design Report. 2005.

[20] ATLAS Collaboration. Performance of the ATLAS silicon pattern recognition al-gorithm in data and simulation at

ps = 7 TeV. (ATLAS-CONF-2010-072), 2010.

[21] RA. Rosenfeld and J. Faltz. Sequential operations in digital picture processing. J.ACM, 13(4):471–494, 1966.

[22] R. Frühwirth. Application of Kalman filtering to track and vertex fitting. Nucl.Instrum. Meth., A262:444–450, 1987.

[23] ATLAS Collaboration. Muon reconstruction performance of the ATLAS detectorin proton–proton collision data at

ps =13 TeV. Eur. Phys. J., C76(5):292, 2016.

[24] D. Wicke. A new algorithm for solving tracking ambiguities. (LC-TOOL-1999-007-TESLA):219–228, 1999.

[25] ATLAS Collaboration. A neural network clustering algorithm for the ATLAS sili-con pixel detector. JINST, 9:P09009, 2014.

[26] R. Jansky. Truth seeded reconstruction for fast simulation in the ATLAS experi-ment. (CERN-THESIS-2013-194), 2013.

[27] ATLAS Collaboration. The optimization of ATLAS track reconstruction in denseenvironments. (ATL-PHYS-PUB-2015-006), 2015.

[28] T. Sjostrand, S. Ask, J. R. Christiansen, R. Corke, N. Desai, et al. An introductionto PYTHIA 8.2. 2014.

[29] ATLAS Collaboration. Summary of ATLAS Pythia 8 tunes. (ATL-PHYS-PUB-2012-003), 2012.

[30] A.D. Martin, W.J. Stirling, R.S. Thorne, and G. Watt. Parton distributions for theLHC. Eur.Phys.J., C63:189–285, 2009.

[31] ATLAS Collaboration. Topological cell clustering in the ATLAS calorimeters andits performance in LHC Run 1. (CERN-PH-EP-2015-304), 2016.

[32] M. Cacciari, G. P. Salam, and G. Soyez. The anti-kT jet clustering algorithm. JHEP,04:063, 2008.

[33] J. M. Butterworth, A. R. Davison, M. Rubin, and G. P. Salam. Jet substructureas a new higgs-search channel at the large hadron collider. Phys. Rev. Lett.,100:242001, Jun 2008.

[34] M. Cacciari, G. P. Salam, and G. Soyez. Fastjet user manual. Eur. Phys. J.,C72:1896, 2012.


[35] ATLAS Collaboration. Jet energy measurement and its systematic uncertainty inproton-proton collisions at

ps = 7 TeV with the ATLAS detector. Eur. Phys. J.,

C75:17, 2015.

[36] D. Krohn, J. Thaler, and LT Wang. Jet Trimming. JHEP, 1002:084, 2010.

[37] ATLAS. Jet energy measurement with the ATLAS detector in proton-proton colli-sions at

ps = 7 TeV. Eur. Phys. J. C, 73:2304, 2013.

[38] ATLAS Collaboration. Performance of jet substructure techniques for large-Rjets in proton–proton collisions at

ps = 7 TeV using the ATLAS detector. JHEP,

1309:076, 2013.

[39] M. Cacciari and G. P. Salam. Pileup subtraction using jet areas. Phys. Lett.,B659:119–126, 2008.

[40] ATLAS Collaboration. Calibration of b-tagging using dileptonic top pair events ina combinatorial likelihood approach with the ATLAS experiment. (ATLAS-CONF-2014-004), 2014.

[41] ATLAS Collaboration. Commissioning of the ATLAS high-performance b-taggingalgorithms in the 7 TeV collision data. (ATLAS-CONF-2011-102), 2011.

[42] G. Apollinari, I. Bejar Alonso, O. Brüning, M. Lamont, and L. Rossi. High-Luminosity Large Hadron Collider (HL-LHC): Preliminary Design Report. 2015.

[43] ATLAS Collaboration. ATLAS Phase-II Upgrade Scoping Document. CERN-LHCC-2015-020. LHCC-G-166, 2015.

[44] ATLAS Collaboration. Expected performance of the ATLAS Inner Tracker at theHigh-Luminosity LHC. (ATL-PHYS-PUB-2016-025), 2016.

[45] R. D. Ball et al. Parton distributions for the LHC Run II. JHEP, 1504:040, 2015.

[46] GEANT4 Collaboration. GEANT4–A Simulation toolkit. Nucl.Instrum.Meth.,A506:250–303, 2003.

[47] ATLAS Collaboration. Measurement of performance of the pixel neural networkclustering algorithm of the ATLAS experiment at

ps = 13 TeV. (ATL-PHYS-PUB-

2015-044), 2015.

[48] ATLAS Collaboration. ATLAS jet energy scale uncertainties using tracks in protonproton collisions at

ps = 7 TeV. (ATLAS-CONF-2011-067), 2011.

[49] T. Gleisberg, S. Höche, F. Krauss, M. Schönherr, S. Schumann, et al. Event gen-eration with SHERPA 1.1. JHEP, 0902:007, 2009.

[50] M. Bahr et al. Herwig++ physics and manual. Eur. Phys. J., C58:639–707, 2008.

[51] ATLAS Collaboration. Jet mass reconstruction with the ATLAS detector in earlyRun 2 data. (ATLAS-CONF-2016-035), 2016.

[52] CMS Collaboration. Particle-flow event reconstruction in CMS and performancefor jets, taus, and MET. (CMS-PAS-PFT-09-001), 2009.


[53] M. Son, C. Spethmann, and B. Tweedie. Diboson-jets and the search for resonantZh production. JHEP, 08:160, 2012.

[54] A. Katz, M. Son, and B. Tweedie. Jet substructure and the search for neutralspin-one resonances in electroweak boson channels. JHEP, 03:011, 2011.

[55] S. Schaetzel and M. Spannowsky. Tagging highly boosted top quarks. Phys. Rev.,D89(1):014007, 2014.

[56] A. J. Larkoski, F. Maltoni, and M. Selvaggi. Tracking down hyper-boosted topquarks. JHEP, 06:032, 2015.

[57] S. Bressler, T. Flacke, Y. Kats, S. J. Lee, and G. Perez. Hadronic calorimetershower size: Challenges and opportunities for jet substructure in the super-boosted regime. Phys. Lett., B756:137–141, 2016.

[58] ATLAS Collaboration. A study of the material in the ATLAS inner detector usingsecondary hadronic interactions. JINST, 7:P01013, 2012.

[59] ATLAS Collaboration. Early inner detector tracking performance in the 2015data at

ps = 13 TeV. (ATL-PHYS-PUB-2015-051), 2015.

[60] ATLAS Collaboration. Alignment of the ATLAS inner detector and its performancein 2012. (ATLAS-CONF-2014-047), 2014.

[61] G. Altarelli, B. Mele, and M. Ruiz-Altaba. Searching for new heavy vector bosonsin pp̄ colliders. Z. Phys. C, 45:109, 1989.

[62] E. Eichten, I. Hinchliffe, K. D. Lane, and C. Quigg. Super collider physics. Rev.Mod. Phys., 56:579–707, 1984.

[63] C. Quigg. Gauge theories of the strong, weak, and electromagnetic interactions.Princeton University Press, page 227, 2013.

[64] J. C. Pati and A. Salam. Lepton number as the fourth color. Phys. Rev. D, 10:275–289, 1974. [Erratum: Phys. Rev.D11,703(1975)].

[65] H. Georgi and S. L. Glashow. Unity of all elementary particle forces. Phys. Rev.Lett., 32:438–441, 1974.

[66] H. Georgi. The state of the art - gauge theories. AIP Conf. Proc., 23:575–582,1975.

[67] H. Fritzsch and P. Minkowski. Unified interactions of leptons and hadrons. AnnalsPhys., 93:193–266, 1975.

[68] L. Randall and R. Sundrum. A Large mass hierarchy from a small extra dimension.Phys. Rev. Lett., 83:3370–3373, 1999.

[69] T. Han, J. D. Lykken, and RJ. Zhang. On Kaluza-Klein states from large extradimensions. Phys. Rev. D, 59:105006, 1999.

[70] K. Agashe, H. Davoudiasl, G. Perez, and A. Soni. Warped gravitons at the LHCand beyond. Phys. Rev. D, 76:036006, 2007.


[71] O. Antipin, D. Atwood, and A. Soni. Search for RS gravitons via WLWL decays.Phys. Lett. B, 666:155–161, 2008.

[72] O. Antipin and A. Soni. Towards establishing the spin of warped gravitons. JHEP,0810:018, 2008.

[73] G. C. Branco, P. M. Ferreira, L. Lavoura, M. N. Rebelo, M. Sher, and J. P. Silva.Theory and phenomenology of two-Higgs-doublet models. Phys. Rept., 516:1–102, 2012.

[74] M. Perelstein. Little higgs models and their phenomenology. Prog. Part. Nucl.Phys., 58(1):247 – 291, 2007.

[75] F. Sannino. Technicolor and beyond: Unification in theory space. J. Phys. Conf.Ser., 259(1):012003, 2010.

[76] E. Eichten and K. Lane. Low-scale technicolor at the Tevatron and LHC. Phys.Lett. B, 669:235–238, 2008.

[77] S. Catterall, L. Del Debbio, J. Giedt, and L. Keegan. MCRG minimal walkingtechnicolor. Phys. Rev. D, 85:094501, 2012.

[78] J.R. Andersen, O. Antipin, G. Azuelos, L. Del Debbio, E. Del Nobile, et al. Discov-ering technicolor. Eur. Phys. J. Plus, 126:81, 2011.

[79] D. Barducci, A. Belyaev, S. Moretti, S. De Curtis, and G. M. Pruna. LHC physics ofextra gauge bosons in the 4D Composite Higgs Model. EPJ Web Conf., 60:20049,2013.

[80] D. Pappadopulo, A. Thamm, R. Torre, and A. Wulzer. Heavy vector triplets: bridg-ing theory and data. JHEP, 2014(9), 2014.

[81] ATLAS Collaboration. Search for high-mass diboson resonances with boson-tagged jets in proton-proton collisions at

ps = 8 TeV with the ATLAS detector.

(CERN-PH-EP-2015-115), 2015.

[82] CMS Collaboration. Search for massive resonances in dijet systems containingjets tagged as W or Z boson decays in pp collisions at

ps= 8 TeV. JHEP, 1408:173,

2014.

[83] CMS Collaboration. Search for massive resonances decaying into pairs of boostedbosons in semi-leptonic final states at

ps = 8 TeV. JHEP, 1408:174, 2014.

[84] ATLAS Collaboration. Search for resonant diboson production in the `′`′qq̄ finalstate in pp collisions at

ps = 8 TeV with the ATLAS detector. Eur. Phys. J. C,

75:69–90, 2015.

[85] ATLAS Collaboration. Search for production of WW /W Z resonances decaying toa lepton, neutrino and jets in pp collisions at

ps = 8 TeV with the ATLAS detector.

Eur. Phys. J., C75(5):209, 2015. [Erratum: Eur. Phys. J.C75,370(2015)].

[86] ATLAS Collaboration. Searches for heavy diboson resonances in pp collisions atps = 13 TeV with the ATLAS detector. JHEP, 09:173, 2016.


[87] CMS Collaboration. Search for massive resonances decaying into pairs of boostedbosons in semi-leptonic final states at

ps = 8 TeV. JHEP, 08:174, 2014.

[88] ATLAS Collaboration. Search for resonances with boson-tagged jets in 15.5 fb−1

of pp collisions atp

s = 13 TeV collected with the ATLAS detector. (ATLAS-CONF-2016-055), 2016.

[89] ATLAS Collaboration. Improved luminosity determination in pp collisions atp

s= 7 TeV using the ATLAS detector at the LHC. Eur. Phys. J. C, 73(8):2518, 2013.

[90] J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O. Mattelaer, H. S. Shao,T. Stelzer, P. Torrielli, and M. Zaro. The automated computation of tree-level andnext-to-leading order differential cross sections, and their matching to partonshower simulations. JHEP, 07:079, 2014.

[91] L. Randall and R. Sundrum. An alternative to compactification. Phys. Rev. Lett.,83:4690–4693, 1999.

[92] ATLAS Collaboration. Electron efficiency measurements with the ATLAS detec-tor using the 2012 LHC proton-proton collision data. (ATLAS-CONF-2014-032),2014.

[93] ATLAS Collaboration. Selection of jets produced in 13 TeV proton-proton colli-sions with the ATLAS detector. (ATLAS-CONF-2015-029), 2015.

[94] ATLAS Collaboration. Identification of boosted, hadronically-decaying W and Zbosons in

ps = 13 TeV Monte Carlo Simulations for ATLAS. (ATL-PHYS-PUB-

2015-033), 2015.

[95] B.R. Webber. Average multiplicities in jets. Phys.Lett., B143:501, 1984.

[96] G. Choudalakis and D. Casadei. Plotting the differences between data and ex-pectation. Eur. Phys. J. Plus, 127(2):25, 2012.

[97] ATLAS Collaboration. Luminosity determination in pp collisions atp

s = 7 TeVusing the ATLAS detector at the LHC. Eur. Phys. J., C71:1630, 2011.

[98] G. Cowan, K. Cranmer, E. Gross, and O. Vitells. Asymptotic formulae forlikelihood-based tests of new physics. Eur.Phys.J.C, 71:1554, 2011.

track reconstruction in dense environments and …with dense environments are produced much more...

Documents